[wave <-> water] support complete roundtripping of attention

[martin-luecke]

The MLIR roundtrip (FX -> Water MLIR -> FX) currently passes all 47 compilation stages for GEMM but fails at every stage for vanilla attention. Running the progressive roundtrip test reveals different distinct root causes that together block all 47 passes. Fixing these will bring attention to full roundtrip parity with GEMM.

An initial investigation shows 4 issues, but more might come up after fixing these:

• Water emitter crashes on iterate blocks whose loop-carried values are reductions #1020 (passes 1-7)
• Water dialect binary ops reject broadcast-compatible operands (rank mismatch) #1021 (passes 8-11)
• FX importer doesn't handle mapping attribute on read/write ops #1022 (passes 12-13, 35-47)
• Water emitter rejects multi-operand (expanded) reductions #1023 (passes 14-34)