Roll back compile cache entry when the first trace throws

[adityasingh2400]

Proposed changes

Fixes #3624.

The per-call lambda built by compile() marks a cache entry non-empty before it traces it:

if (entry.empty) {
  entry.empty = false;                 // marked filled BEFORE the trace
  entry.constants = std::move(constants);
  std::tie(entry.inputs, entry.outputs, entry.extra) =
      compile_trace(fun, inputs, shapeless);   // may throw
  ...
}

If compile_trace (or the subsequent dfs, simplify, or fuse steps) throws on the first trace, the entry is left with empty == false but no inputs, outputs, or tape. A later call with matching inputs then find()s that entry, sees it is not empty, skips tracing, and compile_replaces against an empty tape, returning empty outputs as a spurious success instead of re-tracing or re-raising. A nullary compiled function is the sharpest case since its inputs always match.

This moves entry.empty = false to after the trace, dfs, simplify, and fuse steps complete, so a throwing first trace leaves the entry empty and a later call re-traces cleanly while a failure stays a failure. This is the approach @zcbenz endorsed on the issue. A retry reassigns constants, inputs, outputs, and the tape when it re-enters the fill path, so the partially assigned fields from the failed attempt are overwritten.

Testing

Added a C++ regression test in tests/compile_tests.cpp that compiles a nullary function which raises on its first invocation, then calls it twice. Without the fix the second call returns 0 outputs instead of raising. With the fix both calls raise, and once the function stops raising a retry traces cleanly and returns the correct result.

Verified on macOS (CPU-only build): the new test fails on main (second call did not throw, out.size() was 0) and passes with the fix. The full compile suite (23 cases, 111 assertions) is green. Both touched files pass clang-format.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)

[zcbenz]

I preferred the other fix: #3624 (comment)

[adityasingh2400]

Done, reworked to that approach in 0fde210. The flag move is all that is needed since a failed first trace now leaves the entry empty and the retry path overwrites the partially assigned fields when it re-traces. The regression test is unchanged and still covers the throwing first trace, the throwing retry, and the clean retrace.