Fix workgroup reordering for large shapes

[xintin]

In read_write.py, the split_index function splits an index expression into workgroup-independent (wg) and thread-dependent (th) parts for linearize_memref. The wg part becomes the reinterpret_cast base offset, while the th part becomes the store index.

When reorder_workgroups=True with dynamic dimensions, the output index contains Piecewise(tail_new_wg0, main_new_wg0) expressions.
After substituting WORKGROUP_0/1=0, the difference diff = src - thread_dependent_index still contains dynamic symbols (M, N, K).

The guard detects these residual symbols and falls back:

thread_independent_index = sympy.sympify(0)   # wg offset = 0
thread_dependent_index = src                  # full address as thread index

This makes the full linear address (M_row * N + N_col) the store index, which overflows 32-bit when the output buffer exceeds ~4 GB in bytes (M*N*4 > 2**32).

In the reorder_workgroups=False case, diff = block_id_x * 64 has no residual symbols (block_id_x is a wg symbol), so the split works correctly, embedding the large wg offset into the reinterpret_cast base and keeping only a small thread offset.

[harsh-nod]

Could you add some lit test for this that shows the IR?