[Frontend] Fix conv wrapper crash when the input is a view of a lower-rank buffer

[YWHyuk]

Fixes a conv codegen bug where the generated wrapper crashes with

padded_shape[3] += 2 * 0
IndexError: list index out of range

Root cause

A conv input node can be a ReinterpretView. call_kernel passes template args by
buffer name, so the wrapper is handed the base buffer, whose rank/shape may differ
from the view codegen assumed:

	X
codegen (extract_info)	X.layout.size = [1, 32, 8, 8] (the 4D view)
runtime (wrapper)	X.shape = (1, 64, 32) (the base buffer)

The wrapper then inferred geometry from the runtime tensor (list(X.shape), X.shape[2],
X.shape[3]) and blew up.

It only fires when the view is free: a contiguous (B, N, C) buffer already is the
channels_last layout of (B, C, H, W), so inductor inserts no materializing copy. That is
exactly the transformer (B, N, C) -> (B, C, H, W) pattern — e.g. SegFormer's
efficient-attention spatial-reduction conv (Conv2d(C, C, k=sr, stride=sr), padding 0).
When the layout does not match, inductor inserts a copy kernel, a real 4D buffer arrives,
and the bug is invisible. That is why it went unnoticed.

Fix

Stop inferring geometry from the runtime tensor. Bake the size/stride/offset the codegen
used into the wrapper and rebuild the logical NCHW inputs from it:

X = reinterpret_tensor(X, (1, 32, 8, 8), (0, 1, 256, 32), 0)
W = reinterpret_tensor(W, (16, 32, 2, 2), (128, 1, 64, 32), 0)
X_padding = torch.zeros((1, 32, 8, 8), device=X.device, dtype=X.dtype)
X_padding[:, :, 0:8, 0:8] = X

X.layout describes the data within the storage of the tensor that arrives (both come from
the same node), so the reinterpret is correct — and idempotent — whether the wrapper is
handed the base buffer, an already reinterpreted view, or a materialized copy.

The conv templates were the only place reading .shape at runtime; everything else already
derives geometry from the node layout at codegen time. No change to call_kernel,
mlir_argdefs, the MLIR signature, or arg_attributes — the call site still passes the
raw base buffer and it now works.

Also fixed for free: the padding buffer was allocated via torch.zeros(shape) (always f32)
then .to(device=...) (CPU round-trip). It now allocates directly on device with X.dtype.

Verification

• New regression test tests/ops/conv/test_conv_view_input.py: fails with the above
  IndexError on the pre-fix wrapper, passes after (max diff 2.3e-05 / 2e-04 vs CPU).
• Normal conv path (real 4D buffer, padding=1) unchanged and numerically correct
  (max diff 3.8e-05 vs CPU); reinterpret_tensor is an identity there.

Follow-ups (not in this PR)

• padding == 0 fast path: shapes are literals now, so the zeros alloc + full copy can be
  skipped for 1x1 / strided convs.
• mlir_argdefs derives buffer_types numel from the base buffer, so a slicing view
  feeding e.g. gemm can still get a wrong MLIR arg size. Same class of bug, separate fix.
• def_conv_kernel patches the signature with re.sub(r'(\d+)(?=xf32)', ...), hardcoding
  xf32; an f16 conv would not get its padded input size applied.

Found while checking #254 on this branch: SegFormer no longer hits the originally reported
Not supporting format error, but was blocked here instead.