util, swaps: fix nostr announcement PoW mining

[Amperstrand]

Fixes #10858

Summary

Three defects in the swapserver nostr announcement PoW path, all introduced in 947094c1b (PR #9551):

1. nostr_pow_worker returned the builtin hash instead of the sha256 digest of the winning preimage.
2. gen_nostr_ann_pow collected the result via done.pop() on the unordered set returned by asyncio.wait(FIRST_COMPLETED). Workers that observe the shutdown event return (None, None) at their next 1M-hash block boundary, and their results can be delivered to the event loop before the winner's own result (independent channels: manager socket vs executor result pipe), so the freshly-mined winning nonce could be silently discarded and (None, 0) propagated.
3. set_nostr_proof_of_work stored the returned nonce without checking pow_amount >= SWAPSERVER_POW_TARGET. ConfigVar._set_config_value skips its type-check for None, so the race outcome was persisted — the next publish_offer then raised TypeError: hex(None) outside its try/except, killing run_nostr_server (swapserver stops announcing over nostr until restart, then re-mines from scratch).

Changes (+16/−3 across electrum/util.py and electrum/submarine_swaps.py)

• nostr_pow_worker: return digest, nonce
• gen_nostr_ann_pow: wait for all workers (ALL_COMPLETED), collect the result that carries a nonce, raise if none does
• set_nostr_proof_of_work: raise instead of storing a nonce whose achieved work is below the target

Testing

New tests/test_nostr_pow.py (first commit of this PR; the second commit only changes the two source files):

• The tests intentionally do not run PoW searches — winning nonces are precomputed constants and the process pool / multiprocessing manager are faked — so the suite adds ~0.7 s and no CPU load to CI
• 5 of the 7 default tests fail on unpatched master (each maps to one defect or its failure mode); all 7 pass with this patch
• The race is reproduced with faithful fake workers (test_observer_result_completing_first_is_ignored) — deterministic worst-case interleaving, no flakes across repeated runs
• The 2 passing guards (valid-nonce store, sufficient-pinned-nonce reuse) pin the working behavior; the patch changes nothing about those paths
• Adjacent suites tests/test_util.py + tests/test_submarine_swaps.py: 64/64 green
• TestNostrPowSearch runs actual searches and is skipped by default to keep hash loops off CI; locally with ELECTRUM_TESTS_NOSTR_POW=1 python -m pytest tests/test_nostr_pow.py -v it is 9/9 green (~2 s at a 16-bit target)
• Real-process-pool validation (local): 30/30 runs at target-20 bits return valid nonces with this patch; unpatched master 150/150 valid — the race is real but rare (the winner's result normally leads the observers' by a full 1M-hash block), which the deterministic unit test covers regardless

[Amperstrand]

Full A/B logs for the record: https://gist.github.com/Amperstrand/4b477e0fbc14b22f8601c601f964b210

• unpatched master a94e460b5 + tests commit: 5 failed, 2 passed, 2 skipped (deterministic across 3 runs)
• this branch: 7 passed, 2 skipped (across 5 runs); full default suite pytest tests --ignore=tests/regtest --ignore=tests/qml locally: 1024 passed, 8 skipped, 275 subtests, 0 failures (~4 min)
• the 2 skipped tests (TestNostrPowSearch) run real process-pool searches and are intentionally gated off CI; locally with ELECTRUM_TESTS_NOSTR_POW=1 they pass 9/9 (~2 s at a 16-bit target). Also ran the real pool directly at target-20 bits: 30/30 valid nonces with the patch.

If a different collection strategy is preferred (e.g. return as soon as any nonce-bearing result completes, instead of waiting for all workers — slightly more code, but zero added startup latency), happy to rework.