[Codegen] Add bufferization support for CastToRaggedShapeOp and LinearizeRaggedDimsOp (#24193)

This is PR 4/N of landing block sparse changes that are captured in this
branch :


https://github.com/MaheshRavishankar/iree/tree/users/MaheshRavishankar/blockSparseCommits

Adds `BufferizableOpInterfac`e external models for two tensor-ext ops
that describe ragged shapes:
- `iree_tensor_ext.cast_to_ragged_shape` — reinterprets a source buffer
as a rank-(n+1) ragged buffer by
attaching a RaggedShapeAttr as the memref layout.
-`iree_tensor_ext.linearize_ragged_dims` — the inverse view: collapses
the ragged dimensions of a source
buffer back into a single dense dimension and drops the ragged layout.
  
Both ops are pure views, so the interface implementations:
- Report the operands as reads and never as writes (including
`resultBufferizesToMemoryWrite = false`).
- Set `mustBufferizeInPlace = true` and mark the result as
`BufferRelation::Equivalent` to source, so
bufferization does not introduce a copy.
- For `cast_to_ragged_shape`, the result memref's layout is the ragged
encoding promoted from the tensor
encoding; operand buffers use the identity layout.
- For `linearize_ragged_dims`, both operand and result memrefs use the
identity layout; the result simply
drops the ragged encoding.
`bufferize` replaces the tensor-typed op with an identical memref-typed
version of itself — no data is
moved.