osdtrace: add allocator tracing mode (-A)

[taodd]

Summary

Adds osdtrace -A, an allocator tracing mode: one line per call into the BlueStore free-space allocator (Allocator::allocate() overrides — stupid/bitmap/avl/btree/btree2 and the hybrid wrappers, across the class-name variants used by different releases). BlueFS allocators go through the same method, so RocksDB/BlueFS file growth shows up next to object-data allocations.

osd 2 alloc block      want 0x1000 unit 0x1000 -> 0x1000 exts 1 [0xe1b1000~0x1000] lat 12us comm tp_osd_tp
osd 0 alloc bluefs-wal want 0x1200000 unit 0x100000 hint 0x16300000 -> 0x1200000 exts 1 [0x16300000~0x1200000] lat 9us comm bstore_kv_sync

Per call: allocator instance name, want / unit / hint, bytes actually allocated (SHORT, ENOSPC flagged), returned physical extents (first 6 — a split request is a direct fragmentation signal), latency (lock wait + free-space search) and calling thread. -l <ms> filters on latency; -A -b composes with the BlueStore latency probes.

Rebased onto current main (over #156/#163/#165/#175/#177): the probes now go through ATTACH_LIST (with a ret flag for the shared uretprobe) instead of the pre-refactor ad-hoc attach block.

Commits

1. dwarf_parser: resolve inherited members; skip unresolvable probes — find_class_member() resolves declaration-only base-class DIEs through the type cache and adds the base's offset within the derived class to the member offset (an inherited member previously came back with a base-relative offset). handle_function() warns and drops the probes on that function instead of assert()ing when a varpath doesn't resolve on some build.
2. osdtrace: add allocator tracing mode (-A) — entry uprobes stash the request per {thread, nesting depth}; one shared uretprobe computes latency and walks the returned PExtentVector with _M_start/_M_finish offsets from DWARF (the mempool allocator puts two pointers ahead of them, so they sit at 16/24, not 0/8). Nested calls (hybrid → its internal bitmap) are matched by depth and tagged (nested dN). The allocator name is read via a never-attached varpath on {AllocatorBase,Allocator}::get_name (its SocketHook is only fully defined in that CU) reused at every allocate() entry.
3. osdtrace: look up BPF programs by name, not skeleton index — func_progid mapped functions to a position in the skeleton's progs array; adding a program anywhere but last silently shifted every later index (rebasing the allocator probes onto _txc_add_transaction did exactly that — every allocator id was off by one). Now a function → BPF program name map resolved with bpf_object__find_program_by_name(), plus a check at skeleton open that every mapped name exists.
4. gitignore: ignore generated src/ceph_btf_local.h — Makefile-generated, removed by make clean, like embedded_dwarf_data.h.
5. dwarf: regenerate Ubuntu 20.2.0 osdtrace references (amd64, amd64v3) — exported with osdtrace -j on hosts whose ceph-osd build-ids match the two references (the amd64v3 file is the one the regenerate workflow deliberately skips).
6. dwarf: normalize workflow-written references to the tool's native format — 39 files the old workflow had re-serialized with indent=4, rewritten as the tools emit them (2-space); whitespace only, every file parses to the identical document. One-time consequence of ci: make regenerate-all-dwarf run end-to-end and commit tool output verbatim #180's "commit tool output verbatim".
7. dwarf: regenerate osdtrace references with the allocator probes — output of a Regenerate all DWARF references dispatch on this branch (run 32146786374 — first full end-to-end run of the ci: make regenerate-all-dwarf run end-to-end and commit tool output verbatim #180 workflow: 52/52 jobs; 3 ~cloud0 jobs needed a rerun after a 4 h apt-mirror hang). All 43 osdtrace references now carry the allocate()/get_name entries present in each binary (5 on Octopus, 6 on Quincy/Reef, 9 on Squid+); the 20.2.0 amd64 file it produced is byte-identical to the one exported on the live host in commit 5. Radostrace references are unchanged and no existing value moved anywhere.

Compatibility

• Default/-s/-b modes are unchanged; existing DWARF JSONs and embedded data keep working for them.
• Every embedded reference now includes the allocator probes, so -A works out of the box; with an older user-supplied JSON (-i) it fails to attach with a hint (-A needs DWARF data that includes the allocator probes … regenerate with -j).

Verification (Ceph 20.2.0, Ubuntu 26.04 VM, live osd.2 with debug symbols)

• osdtrace -j on the live OSD exports all 9 allocator functions present in that binary plus the get_name name donor; _M_start/_M_finish resolve at 16/24 (inherited-member path).
• osdtrace -A -i <that json> under rados bench 4K writes: 16 probes attached (8 entry + 8 return), 10,747 alloc events for 10,739 writes, sequential 4K extents from block, latency p50 19 µs / p99 358 µs / max 3.7 ms; BlueFS calls appear from bstore_kv_sync.
• -A -b -l 1: alloc lines all ≥ 1 ms (20) alongside 40 k bluestore lines; -A against the embedded reference prints the hint and exits.
• Default mode regression: embedded DWARF path, 15 probes attached, 7,200 op rows, no warnings.
• Final binary (regenerated references embedded): osdtrace -A without -i on the same OSD — embedded match, 16 probes attached, 11,729 alloc events for 11,721 writes.
• make -j all warning-free, make test, tox -e test pass; clang-tidy runs in CI.