Fix bounds expressions to respect workgroup reordering

[willghatch]

Fix bounds expressions to respect workgroup reordering

generate_bounds_exprs was not aware of ReorderingConstraints, so when a bound expression contained a raw workgroup symbol (e.g. WORKGROUP_1) it was emitted as-is, mapping directly to block_id_y.
With workgroup reordering active, the actual tile position depends on both block_id_x and block_id_y (via the flattened/swizzled index), so the raw symbol gives an incorrect mask.
This happens when WaveConstraint.get_index_bound() fires -- i.e. when the wave tile size is not divisible by the MMA vector shape (e.g. BLOCK_N=224 with 4 waves gives wave_tile=56, and 56 % 16 != 0).
The returned bound WORKGROUP_1 * BLOCK_N + wave_id * wave_tile + wave_tile contains WORKGROUP_1.
The fix passes reordering_constraints into generate_bounds_exprs and substitutes each workgroup symbol with its reordered expression before storing bounds on the node.
This ensures the emitted affine.apply for masking references both block IDs when reordering is active.

Fixes the 256x224x256 block size for dynamic-shape GEMM. (More specifically the 7.1 dynamic preshuffle with LLVM backend.)

[willghatch]

Well, the top commit that is mine is the thing that this is about. I assume that as the other PRs get merged I'll rebase this until it is actually just one commit. But for now if you want to view, use the commit view.

[harsh-nod]

lgtm!