Compile time of large kernels #322

A first streamlines call on the AhmedML volume (231 seeds, CPU) takes 3.98 s with the disk cache off and 0.46 s warm; the tracing is 0.1 s. _advect is 3.75 s of it: 47,849 IR nodes once its device classes and functions are inlined (526 inlined calls, 208 template methods). This matters for first runs, CI, a new Tack version and MPI jobs.

Inside _advect's 3.75 s: lowering 0.75 s, copy propagation 0.73 s, verify_ir (six full passes) 0.6 s, LLVM IR generation 0.33 s, LLVM parse/optimize/JIT 0.26 s; 29% of the time is walk_ir, a recursive generator called 1.28M times.

• Drop branches that compile-time constants decide, during lowering (13 such branches hold about 14,000 of the 47,849 nodes).
• Verify less outside debug mode: one or two of the six full verifications.
• One-pass copy propagation: one substitution map in one walk.
• A faster IR walker: an explicit stack instead of a recursive generator.

Target: a cold first call of about 1.5 s here. Measure on the GPUs too. #298 (template methods: fewer copied parameters, methods resolved on demand) takes part of this.