Dynamic shape on the C++ trace path (elementwise add)

[YWHyuk]

Summary

Adds torch.compile(dynamic=True) support on the C++ trace path. One compiled
trace producer .so serves any input size: the producer reads its loop bound
from shape_args at runtime, while the per-tile cost table (sampled once, since a
tile's cost is shape-invariant) keys the timing.

Verified end to end: a single dynamic a + b runs at N=1024 (2 tiles, 183
cycles) and N=2048 (4 tiles, 261 cycles) from the same .so, the trace
scaling with shape_args.

What each commit does

1. Guard MLIR tile sizing against symbolic dims — neutralize the tile-fit
   heuristics for a sympy dim (they only minimize a known dim's tail); all gated
   on isinstance(.., sympy.Expr), so static is unchanged.
2. Emit symbolic loop bounds and dynamic memref dims — size symbol becomes a
   scalar-int kernel arg; memref<?xf32>; affine.for ... to %<name>_bound with a
   top-of-function memref.load + index_cast prologue.
3. Make the kernel meta import-safe — stringify sympy in arg_attributes so the
   generated wrapper imports (the extent already arrives as a runtime arg).
4. Skip compile-time Spike validation for dynamic — its fixed-shape validation
   binary can't be instantiated for a runtime extent.
5. Sample per-tile cycles on a one-tile copy — pin_loops_to_one_tile (general
   for static + dynamic) runs gem5 sampling on a one-tile copy; the symbolic IR is
   kept for the producer / cost table.
6. Emit a dynamic-shape trace producer — build_skeleton skips the loop_end
   serialization it never needed; lower_to_emitc re-sources each still-used size
   arg from shape_args[k] via emitc.subscript, so the producer loop reads
   for (iv=0; iv<shape_args[k]; iv+=step).
7. Pass the runtime shape via the attribute file — the scalar size goes into the
   existing per-kernel attribute YAML (shape_args), run_standalone passes
   --attribute, and main.cc fills shape_args for run_producer.
8. Test — tests/ops/elementwise/test_dynamic_add.py.

Known limitations / follow-ups

• Timing only: the Spike functional run (which fills the output tensor) is
  skipped for dynamic, so output values are zeros today; the test checks output
  shape. Functional output for dynamic is a follow-up.
• Static cost sampling can use the (already general) pin_loops_to_one_tile once it
  is decoupled from run_tog; op coverage beyond contiguous 1D add (matmul /

  1D-dynamic: symbolic-stride addressing, multi-symbol shape_args) is future work.

• One change to the loop-padding mlir-opt pass (skip a symbolic bound) lives in the
  LLVM fork and is not part of this PR.

Base: stacked on feature/togsim-cpp-trace (#267).

🤖 Generated with Claude Code