Allow typecast fusion

[vtsilytskyiTT]

Problem description

ttl.typecast could not participate in fused ttl.compute regions, which forced dtype-changing expressions to break fusion or fail during TTL-to-compute lowering. This also made mixed dtype fusion behavior unclear, especially when f32 inputs required unpack_to_dest_fp32 handling.

What's changed

This change enables ttl.typecast fusion by deriving tile result types from each source op's tensor result instead of relying on a single default tile type. It updates compute lowering, Python AST/API type handling, and kernel config assignment so fused compute regions can preserve dtype-changing intermediates correctly.
The compute kernel config logic now tracks unpack_to_dest_fp32 per CB and detects conflicts where the same f32 CB is consumed by both FPU and SFPU strategies. Tests were added and updated to cover typecast fusion, mixed dtype rejection, and unpack_to_dest_fp32 positive, negative, and conflict cases.

Ticket

#264

Checklist

• New/Existing tests provide coverage for changes

[vtsilytskyiTT]

Resolve conflicts.

[zoecarver]

Very cool! LGTM. Thank you, Vlad!

[brnorris03]

This still uses outputTileType, which is the final fused sink type, not necessarily this particular op's result type. That breaks cases like typecast(bcast(...)), where the bcast result should keep its own dtype and the later ttl.tile_typecast should perform the conversion.

The same applies to the matmul special cases below. These special-case emitters should derive the result tile type from the current source op, like emitTileOpFor() does.

[brnorris03]

Should this be a pass error instead of easy to miss warning (since could lead to losing f32 precision)?

[brnorris03]

Can you maybe add a couple fused testcases like typecast(matmul) or typecast(matmul + a)?