feat(vpto): add L1 fill intrinsic support

[erhsh]

Motivation

This PR adds a low-level PTO/VPTO operation for filling an L1 (MAT/CBUF) region with a repeated 16-bit or 32-bit value. The intent is to expose the hardware create_cbuf_matrix capability directly so that callers can initialize L1 without first materializing data in GM or UB and issuing a separate copy.

What changes

• Adds pto.mte_fill_l1 to the PTO dialect and documents its semantics.
• Adds verifier checks for the L1 destination, fill word width, byte offset, and constant configuration fields.
• Lowers the operation in both VPTO LLVM emitter paths to:
  • llvm.hivm.CREATE.CBUF.MATRIX.v3.u16
  • llvm.hivm.CREATE.CBUF.MATRIX.v3.u32
• Exposes the operation through PTODSL as pto.mte_fill_l1(...).
• Adds lowering, verifier, PTODSL surface, and public-export coverage.

Parameter mapping

The hardware configuration operand is assembled as:

config = repeat_times
       | (block_num_32b << 16)
       | (dst_gap_32b << 32)

Each field is 15 bits. byte_offset is applied to the destination pointer before calling the intrinsic, and the low 16 or 32 bits of raw_value provide the repeated fill pattern.

Semantically, each repeat fills block_num_32b consecutive 32-byte blocks, skips dst_gap_32b blocks, and then starts the next repeat.

Reviewer notes

The main points worth checking are:

1. Whether the intrinsic names and argument order match the CANN 9.0 and default VPTO emitter conventions.
2. Whether the packed field positions and 15-bit masks match create_cbuf_matrix.
3. Whether the exposed PTODSL parameter names make the hardware units (bytes vs. 32-byte blocks) sufficiently clear.