Treat a process that exits during Identify as gone, not unreadable

[ppXD]

Summary

• NativeProcess.Identify still escaped raw exceptions for a process that exits after its /proc/<pid>/stat read: Process.GetProcessById throws ArgumentException ("not running") and Process.StartTime throws Win32Exception ("may have exited or may be privileged"). Treat a process that vanishes mid-read as gone, not unreadable #2053 mapped the IOException from the stat read only. When kill(pid, 0) answers ESRCH, both now become the same IOException("Cannot identify a terminated native process.", inner) the stat read, the Z/X states and the macOS branch already give; otherwise they propagate, so unknowable stays unknowable.
• TreatsAsGone (internal) accepts ArgumentException and Win32Exception alongside IOException, on the same evidence and never the exception's type or message.
• Existing callers are unchanged. IsRunning, IsAlive and KillSession reach TreatsAsGone only from catch (IOException) clauses, so the filter is never handed the two new types there; each already swallows ArgumentException unconditionally and guards Win32Exception with its own IsAbsent probe. Only the new catch (Exception) in Identify sees them.

Test plan

• Unit: NativeProcessGoneTests 10 -> 17 tests, all green (macOS). New: ArgumentException and Win32Exception with an absent pid read as gone; all of IOException, ArgumentException and Win32Exception with a live pid do not; InvalidDataException and UnauthorizedAccessException still do not. Two more feed the runtime's own exceptions rather than built ones: the ArgumentException a real GetProcessById of a reaped pid raises, and (Linux only) the Win32Exception a real StartTime raises for a process that exited after it was opened.
• Mutation: reverting the widening (failure is IOException && IsAbsent(pid)) turns 3 tests red on macOS (the two new rows and the real GetProcessById exception) and 4 on Linux (plus the real StartTime exception); restored by copy and cmp-verified.
• Linux (aarch64 container, no network): the same 17 tests pass, including the Linux-only one. A race harness calling Identify from 4-16 threads while the target process exits and is reaped saw 649-1258 escapes per run over six runs on main (only ArgumentException and Win32Exception) and none over five runs with this change (8x800 three times, 16x400, 4x1500; 200k-490k identities and 200k-440k typed refusals per run).
• Neighbours: LocalProcessDurableRunnerTests 154/154, NativeLaunchRegistryTests 91/91, ProcessLivenessDriftTests 36/36 (integration project, it regex-parses NativeProcess.cs).