[compile] Add dump_mlir_dir option to compile_tile

[ravil-mobile]

Allow callers to specify a MLIR dump directory directly, overriding CUDA_TILE_DUMP_TILEIR.

Description

This turns MLIR dumping from a global, process-level debug switch into a per-call, programmatic option. That matters most for callers that compile many variants of the same kernel in one process and want fine-grained control over which one(s) get dumped and where - the clearest example is autotuning:

• An autotuner (experimental/tile_experimental, or compilation/_export.py's multi-signature export path) may call compile_tile repeatedly for different tile-size/config candidates of the same kernel.
• With only the env var, you couldn't isolate the MLIR for one specific candidate config, or route different configs to different directories, without restarting the process with a different env var each time.
• With dump_mlir_dir, calling code can decide at runtime - e.g. "dump the MLIR for the config that autotuning picked as the winner" or "dump each candidate into its own subdirectory keyed by config" - by just passing the argument on that specific compile_tile() invocation, no environment mutation needed.

Checklist

• [ x] I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.