Add accuracy-routing gate for QNN EP coverage runs

[yuhuchua-qti]

Summary

Replaces the unconditional QnnUnit_Accuracy_* full run in the coverage script's accuracy phase with a gate that skips accuracy cases whose paired snapshot already passed at a matching golden version, and runs the rest.

Skip a case only when both hold:

• its paired snapshot passed (graph structure unchanged), AND
• the golden manifest's QAIRT backend and ORT runtime versions both equal the current ones (either up-level can shift numerics even when the emitted structure is byte-identical).

Any other situation runs accuracy: snapshot drift (failed), golden absent / backend unavailable (skipped, no PASS proof), no paired snapshot (unmapped), or version mismatch / no manifest. Reasons are collected as a union with no short-circuit, so drift is always surfaced in the summary even when a version mismatch already forced a full run.

What's here

• qcom/scripts/linux/accuracy_gate.py — pure decision functions + a thin CLI.
• qcom/scripts/linux/generate_coverage.sh — wires the gate between the snapshot and accuracy phases; prints a developer-facing summary. Any gate failure (missing snapshot JSON, list failure, gate error, empty filter) falls back to a full run, so coverage is never silently dropped.
• qcom/scripts/linux/tests/{conftest.py,test_accuracy_gate.py} — fixture unit tests (synthetic gtest JSON + accuracy lists; no binary, no build).

Stacked on #<PR1: dev/yuhuchua/migrate-clip-ut>.

Fix included

The snapshot→accuracy suite-name pairing regex required an underscore on both sides of the tier token (Snapshot/SessionSnapshot), so bare (no-variant) op-first suites like QnnUnit_Clip_SnapshotTest never matched and fell through as unmapped — forcing those accuracy cases to always run instead of being correctly paired against the snapshot result. Fixed derive_accuracy_suite's regex to not require a trailing underscore, with a regression test covering both the bare and variant suite forms. Confirmed via a real generate_coverage.sh run: unmapped went from 5 to 0 for Clip.

Test plan

• pytest qcom/scripts/linux/tests — 30 passed
• python qcom/build_and_test.py lint — clean
• End-to-end CLI against the local golden cache (manifest with both qairt+ort) + real 14-case accuracy list:
  • both versions match -> run 0, skip 14
  • ORT mismatch -> full run (14)
  • QAIRT mismatch -> full run (14)
• Full coverage_linux_x86_64 build + run on this branch (rebased directly onto PR1's latest tip): component 2504 passed / 12 skipped (env-limited), accuracy 16 passed / 0 failed, gate summary shows unmapped=0 (no manifest locally -> full run, as expected: absent=16 version-mismatch=16)