Fix race in SimpleFixedSizeExemplarReservoir offer/collect (#5431)

[feiiiiii5]

Fixes #5431.

Problem

SimpleFixedSizeExemplarReservoir._find_bucket_index does:

self._measurements_seen += 1                                    # (a)
if self._measurements_seen < self._size:
    return self._measurements_seen - 1
index = randrange(0, self._measurements_seen)                   # (b)

Between (a) and (b), a concurrent collect() can run _reset() and zero
_measurements_seen, producing:

ValueError: empty range for randrange() (0, 0, 0)

FixedSizeExemplarReservoirABC.collect() has a matching race: it iterates
self._reservoir_storage.items() (a defaultdict) while offer() can
insert a new bucket, raising RuntimeError: dictionary changed size during iteration.

Fix

Add a threading.Lock on FixedSizeExemplarReservoirABC.__init__ and
guard offer() and collect() with it.

AlignedHistogramBucketExemplarReservoir overrides offer() (because
it computes its bucket index from the value, not the reservoir state),
so its override must also acquire the lock — otherwise the shared
_reservoir_storage would still be unsafe.

Verification

• New regression test test_concurrent_offer_and_collect_does_not_raise
  runs 2 offer threads + 1 collect thread for 1.0s on a size=1
  reservoir (maximizes the probability of hitting the randrange
  branch on every offer after the first).
• Red → green: with the fix stashed, the test fails with the exact
  `ValueError: empty range for randrange() (0, 0, 0)` from the issue.
  With the fix, it passes.
• Full exemplar test suite passes:

  uv run --python 3.10 --frozen --all-packages python -m unittest \\
    tests.metrics.test_aggregation \\
    tests.metrics.test_exemplarfilter \\
    tests.metrics.test_exemplarreservoir \\
    tests.metrics.integration_test.test_exemplars
  

  → 51 tests pass.
• `ruff check` clean.

Notes

• The lock is on FixedSizeExemplarReservoirABC, not on the base
  ExemplarReservoir, because the base has no shared mutable state to
  protect — the race is specific to the _reservoir_storage defaultdict
  and the subclass-specific reservoir state (_measurements_seen).
• Performance: offer() is now O(1) under the lock (no I/O, no
  allocations beyond the existing defaultdict lookup), so contention
  is bounded. For high-throughput metrics pipelines, the next step would
  be a per-bucket lock or a lock-free reservoir, but neither is in scope
  for this bug fix.

[herin049]

Closing this given that there is already an open PR #5437