bench: accept wire-rounded probability sums per label count

[hsliuustc0106]

Purpose

Fix the live validator bug in benchmarks/bench.py found in #40's first GPU run
(comment):
the probability sum check used a flat math.isclose(..., abs_tol=1e-4), but workers round
probabilities to four decimal places on the wire, so a rounded distribution can legitimately
miss 1 by more than 1e-4. Two MMLU responses summing to 0.9998999999999999 were rejected at
the floating-point boundary, which stopped the first measured GPU run after 498/500 requests.

The tolerance is now derived from the wire format instead of being a flat constant: each of the
n rounded values carries up to half a decimal step (5e-5), so the acceptance bound is
n * 5e-5 (plus epsilon for float representation), and the sum is computed with math.fsum
to remove addition-order noise. A four-label distribution summing to 0.9999 is accepted; a
distribution off by a centi-level (e.g. 0.99) is still rejected, so this widens acceptance
only by what the rounding itself can produce.

The score-vs-expected-value check keeps its abs_tol=1e-4: no failure of that check has been
observed, and its deviation sources (worker-side rounding of a derived scalar) are different.

Test Plan

System1-Omni Version / Commit: main at 69c3485 (after #67)

• New regression test replays the two exact vectors observed in the GPU run
  ([0.1099, 0.1211, 0.5589, 0.2100], [0.1130, 0.5135, 0.3151, 0.0583]) through
  read_answers, and asserts a four-value distribution summing to 0.99 is still rejected.
• Verified the regression test fails on the pre-fix code with the original
  ValueError: probabilities must sum to one and passes with the fix.
• python -m unittest discover -s tests/benchmarks -p 'test_*.py' — 8 tests OK
  (Python 3.13, httpx 0.28.1; CI's 3.11 is equivalent here).
• python benchmarks/bench.py validate benchmarks/smoke.jsonl — OK, unchanged SHA256.

Test Result

All weight-free checks green locally on macOS (arm64). No benchmark numbers are claimed by
this PR; it changes validation only, and it un-blocks judged parity runs for the Laya native
engine chain by no longer rejecting legitimate rounded responses.

Self-review

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.