[Codegen] Materialize encoding info for convolution with NCHWc layout (#24714)

Implements convolution encoding materialization for data-tiled layouts.

Layouts:
- Input:  [N, IC/c0, H, W, c0]
- Filter: [OC/k0, IC/c0, FH, FW, c0, k0]
- Output: [N, OC/k0, OH, OW, k0]

where k0 and c0 are the inner tile sizes for the output and input
channels, respectively.

Scope:

- Only 2D convolutions receive a data-tiled layout. 
- Grouped, depthwise, 1D, and 3D convolutions, and non-unit dilations,
all materialize to the identity layout (existing un-tiled path, no
behavior change for these cases).
   - Arbitrary strides are supported.
- Supports packed rank-5 input/output and rank-6 filter tensors. Rank-4
input/output is also accepted when `N=1` batch dimension has been folded
by upstream passes.
- Currently supports fp32 convolutions only.

Supported input forms:
- nhwc_hwcf
- nchw_fchw
- nhwc_fhwc

All are materialized to the same packed NCHWc layout above.

Architecture-specific inner tiles:

| Architecture | Condition | Inner tile (OC, IC) | Target ISA |

|--------------|-------------------|---------------------|-----------------|
| x86-64 | float, `+avx512f` | `{16,16}` | `VFMADD*` (zmm) |
| arm64 | float | `{8,8}` | `FMLA` (NEON) |

All other configurations materialize to the identity layout.

Layout motivation can be found here:
https://hackmd.io/@phemashekar/conv-dt-layout

---

Co-authored-by: Jelle Schuhmacher <schuehmacher@roofline.ai>
Co-authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: Pooja Hemashekar <hemashekar@roofline.ai>

---------

Signed-off-by: Pooja Hemashekar <hemashekar@roofline.ai>
Co-authored-by: Jelle Schuhmacher <schuehmacher@roofline.ai>
Co-authored-by: Ege Beysel <beyselege@gmail.com>