Add tensor-based TileArray programming and GEMM lowering

[ShangkunLi]

Summary

This PR adds tensor memory programming and standalone lowering for TileArray programs.

It introduces:

• typed tensor arguments and tensor slices;
• configured and dynamic load and store operations;
• TileArray operations with multiple results;
• lowering from a TileArray program to a Taskflow task and Neura kernel;
• a 3×3 weight-stationary GEMM example;
• tests for memory access, lowering, and complete GEMM compilation.

Memory access

Loads and stores can use a configured tensor slice:

value = synl.load(A[:, k], tile=array[0, y])
synl.store(value, target=C[:, x], tile=array[x + 1, 0])

They can also use an explicit runtime address:

value = synl.load(addr=address, dtype=synl.i32, tile=array[0, 1])
synl.store(value, addr=address, tile=array[1, 0])

Example: 3×3 GEMM on a 4×4 TileArray

The west column loads activations, the inner 3×3 Tiles perform MACs, and the south row stores the result:

LD → MAC → MAC → MAC
       ↓     ↓     ↓
LD → MAC → MAC → MAC
       ↓     ↓     ↓
LD → MAC → MAC → MAC
       ↓     ↓     ↓
      ST    ST    ST

Each MAC produces a partial sum flowing south and forwards its activation east:

accumulated, flowing = synl.mac(
    flowing,
    previous_sum,
    stationary=B[k, x - 1],
    tile=array[x, y],
)

Synapse records this program as a TileArrayProgram, wraps it in a Taskflow task, and lowers it to a Neura kernel for template mapping.