[MLIR][Vector] Relax shape_cast unrolling to per-reassociation-group contiguity (#205684)

UnrollShapeCastPattern previously only unrolled a vector.shape_cast when
the target unroll tile was contiguous in the whole result vector
(isContiguous(targetShape, resultShape)). This rejected valid cases such
as:
```mlir
  %0 = vector.shape_cast %src : vector<8x32xf8> to vector<8x1x32xf8>  // target tile [8, 1, 4]
  %1 = vector.shape_cast %src : vector<8x32x32xf8> to vector<256x32xf8>  // target tile [16, 4]
```
This PR factors the source and result shapes of a shape_cast into
independent reassociation groups — maximal aligned source/result dim
ranges that hold equal element counts:
```
  - 8x32 → 8x1x32 ⇒ groups {8 ↔ 8x1}, {32 ↔ 32}
  - 8x32x32 → 256x32 ⇒ groups {8x32 ↔ 256}, {32 ↔ 32}
```
Instead of requiring the target tile to be contiguous in the whole
result vector, it only checks that the tile is contiguous within each
group.
This criterion is correct because a shape_cast preserves row-major
linear order. The groups partition the dimensions into blocks whose
element ranges don't overlap, so the global linear index breaks down
into one independent sub-index per group. A tile that is contiguous
within each group therefore corresponds to a slice that is contiguous
within each group of both the source and result.
  
     Assist-by-Claude

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>