[S-TIR][MetaSchedule] Make evolutionary search resilient to trace replay failures

[cchung100m]

Hi Committers,

This PR is trying to fix issues #17934. Any suggestions would be appreciated if you are available.

Root Cause

• During EvolutionarySearch candidate generation, trace->ApplyToSchedule(...) could throw ScheduleError.
• The exception was propagated through parallel execution and aborted tuning.
• Error handling was inconsistent between measured and unmeasured paths, and failure visibility was limited.

Solutions

• Catch trace replay failures in ThreadedTraceApply::Apply and return nullopt instead of crashing.
• Add trace replay failure counting (trace_fail_counter_) and accessor (TraceFailCount()).
• Align measured path PickBestFromDatabase with unmeasured behavior: skip invalid candidates and continue.
• Add visible WARNING logs when trace replay failures occur (to avoid silent failures).

[gemini-code-assist]

Code Review

This pull request enhances the robustness of the evolutionary search strategy by gracefully handling trace replay failures. Key changes include wrapping schedule application in try-catch blocks within ThreadedTraceApply, introducing an atomic counter to track these failures, and updating PickBestFromDatabase and SampleInitPopulation to log warnings and filter out invalid schedules instead of terminating. The review feedback suggests replacing DLOG with TVM_PY_LOG to ensure that detailed failure information is visible in production builds as well as debug builds.

[cchung100m]

Hi @tlopex @mshr-h

This PR is trying to fix issues #17934. Any suggestions would be appreciated if you are available.

[tlopex]

Overall LGTM. Could you have a look at the review of Gemini and fix them as Gemini suggested?

[cchung100m]

Hi @tlopex
Thanks for the prompt reply. I updated the part you mentioned. 😄

[cchung100m]

Thanks to @tlopex 😄