Fix conv transpose lowering

[jakesabathia2]

Converting a model with a squeeze/reshape after a transposed convolution (nn.ConvTranspose1d/2d) failed MLIR verification at save_asset() — even with fully static input shapes:

  %22 = "coreai.shrink_dims"(%20, %21) : (tensor<?x?x?xf32>, tensor<1xsi32>) -> tensor<?x?xf32>
  RuntimeError: failed to persist mlasset

This blocked converting audio source-separation models (e.g. Meta's htdemucs, whose iSTFT/decoder uses transposed convolutions heavily). This PR fixes the transposed-conv lowering so static shapes are preserved end-to-end.

Root cause

Not the transposed conv op itself (its output is static). Two issues in _conv_transpose
(coreai_torch/_aten_to_core.py):

1. Static-shape erasure (the reported bug). The 1D path runs conv_transpose2d on an expanded [N,C,W,1] tensor
   and then reshaped back to [N,C,W] using a runtime shape tensor:
   coreai.reshape(result, coreai.slice_(coreai.get_shape(result), [0], [3], [1]))
2. reshape with a runtime shape operand can't infer a static result type, so it produced tensor even
   when the conv output was fully static. A downstream squeeze (coreai.shrink_dims) then saw dynamic dims and the
   verifier rejected it.
3. Incorrect output_padding (latent, surfaced while fixing externalize: SDPA submodule re-export drops the upper bound on the key-length dim with a static query + dynamic KV context #1). output_padding was emulated by pre-padding the
   input, which in a stride>1 transposed conv adds stride to the output instead of output_padding (e.g. padding=0,
   output_padding=1, stride=2 → output length 136 vs PyTorch's 135). This was silently wrong at runtime and only
   masked at compile time by the dynamic result type.

Fix

• Shape-preserving shrink-back (1D): replace the runtime-shape reshape with coreai.shrink_dims(result, [-1])
  (the true inverse of the expand_dims used to go 1D→2D) when the trailing dim is statically 1; fall back to the
  runtime reshape only when the input is dynamic (where the op reports all dims as dynamic).
• Native output_padding: the Core AI conv_transpose2d op's native padding and output_pad args already match
  PyTorch semantics exactly, so the entire manual pre-pad/crop juggling was removed and padding/output_padding
  are passed straight through. This is both correct and simpler (−55/+19 lines).

Tests

• New regression test tests/ops/test_ops.py::test_conv_transpose_static_shape_squeeze — squeeze/reshape after
  conv_transpose1d/2d with static input, output_padding ∈ {0, 1}; validates numerics against torch eager and that
  save_asset() passes MLIR verification.
• Updated tests/ops/test_ops_ir.py::TestConvolutionIR::test_conv_transpose2d FileCheck to the new (simpler,
  correct) IR — the op now emits 1x8x8x8 directly with no pad/slice.
• Existing test_conv_transpose (18 cases, incl. output_padding and dynamic shapes, 1D/2D) continues to pass.

Verification

• Radar repro (ConvTranspose1d(4,1,8,stride=2) + .squeeze(1), static input): converts and save_asset()
  succeeds, no dynamic dims in IR.
• output_padding=1, padding=0 (the previously-wrong case): now numerically correct and static.
• Dynamic-input transposed conv: unchanged (still uses the runtime-reshape fallback).
• 65/65 conv tests pass; ruff clean. Full suite shows no regressions attributable to this change