[MicroPerf] Four profiler-guided hot-path wins from self-build tracing

[T-Gro]

Four independent, output-identical hot-path optimizations found by profiling the compiler compiling itself (dotnet-trace sampled thread-time plus targeted microbenchmarks). Emitted IL and metadata are byte-for-byte unchanged on every workload checked.

• Metadata table sort: replaced the array → list → List.sortWith → array round-trip with a packed int64 key [Val:32][Tag:8][pos:24] sorted by Array.Sort and gathered. Isolated microbench: 4–8× faster, ~6× less allocation (65,536 rows: 10.6 ms → 2.5 ms, 6.55 MB → 1.05 MB; 64 rows: 2.99 µs → 0.43 µs).
• Decision-tree target emit: the postponed-target set was re-sorted by index although it already arrives ordered from an IntMap; it is now folded out directly. That sort's projection was the second-hottest managed method in the self-build trace at 4.09 s; isolated, the rewrite is 2.6–7× faster at ~half the allocation.
• Nullness application: addNullnessToTy now returns the original TType when combining nullness leaves it unchanged, instead of rebuilding a structurally identical node. ~153,000 redundant TType allocations removed per 65k-LOC compile (~5 MB), and the function's own CPU drops 27% (0.52 s → 0.38 s).
• Type-hierarchy walk: hoisted the repeated tryTcrefOfAppTy out of the per-iteration guards (common-subexpression elimination), cutting ~25% off the loop's exclusive CPU in the self-build trace (214 ms → 160 ms).

[github-actions]

⚠️ Release notes required, but author opted out

Warning

Author opted out of release notes, check is disabled for this pull request.
cc @dotnet/fsharp-team-msft