[FEAT] Add a compiled mode to the sanitizer

[mark14wu]

Summary

Adds a compiled mode to the sanitizer: Sanitizer(compile=True) or tile-sanitizer --compile script.py. Each launch is checked statically on the host-compiled TTIR instead of being interpreted. The kernel is not launched, so no GPU is needed and output tensors are not written.

Stacked on #479 (base branch ir-mode-core); review and merge that one first.

Behaviour

• Every autotune config of a launch is checked with Z3 against that launch's scalar arguments, grid and tensor layouts. A launch is ok only if every config is proven in bounds for those arguments.
• Findings, each with a witness (program ids, lanes, loop iteration):
  • out-of-bounds, against the view's element footprint (gaps of strided views count, as in eager mode);
  • integer-overflow of fixed-width address, mask, branch and loop values;
  • division-by-zero.
• Whatever cannot be modeled is unsupported with a typed refusal kind; a solver unknown / timeout never counts as a proof, and each query has its own timeout.
• A kernel that does not compile for the target (e.g. num_ctas>1 on the default cuda:89) is unsupported and the program continues; a call that does not bind the kernel's signature raises exactly as it does untraced.
• Each launch records an IRVerdict with per-config verdicts in Launch.records; tilelens.save() round-trips it.

Evaluation

Differential corpus of 223 kernels against a concrete oracle (Triton's interpreter checking every access): 0 false proofs and 0 false out-of-bounds findings on Triton 3.6 and 3.8. 25 cases (~11%) are unsupported, mostly data-dependent masks, early returns / scf.while, gathers and multiple loops.

Testing

Full relevant set, CPU only: Triton 3.6 1379 passed / 31 skipped; Triton 3.8 1395 passed / 23 skipped. Only the pre-existing failures listed in the base PR remain.

Known limitations

• ok holds for the launch's own arguments; since kernels are not launched, a later launch that reads an earlier launch's outputs sees unwritten values.
• One loop per kernel; scf.while, early returns, calls, gathers and data-dependent masks/branches are reported as unsupported.
• The default target is cuda:89; kernels that need another target can pass target= or set TILELENS_IR_TARGET.

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.105s	0.106s	+1.1%	20 / 20
gemm_oob	0.117s	0.118s	+0.6%	20 / 20
indirect_load	0.022s	0.022s	+0.2%	20 / 20
nested_loop	0.235s	0.238s	+1.3%	20 / 20
block_pointer_loop_advance	0.126s	0.127s	+1.2%	20 / 20
liger_jsd	0.139s	0.140s	+1.2%	20 / 20
flaggems_layernorm	0.393s	0.399s	+1.4%	20 / 20
swiglu	0.172s	0.173s	+0.6%	20 / 20
cross_entropy	0.982s	0.987s	+0.6%	20 / 20
fused_linear_jsd	0.210s	0.213s	+1.4%	20 / 20
Total	2.500s	2.523s	+0.9%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.