Floating-point support (f32/f64 types and operations)

[tetsuo-cpp]

Summary

Add floating-point type support. Currently everything is i64-only, but real GPU kernels (especially matmul) operate on f32/f64.

What's needed

• Float literals (or int-to-float conversion words)
• Float arithmetic: F+ F- F* F/
• Float load/store: F@ F!
• Type conversion: S>F (int to float), F>S (float to int)

Design considerations

The Forth stack is currently untyped (!forth.stack backed by memref<256xi64>). Options for float support:

1. Bit-casting approach: Store f32/f64 as i64 bit patterns on the stack, with F-prefixed ops doing reinterpret casts before/after arithmetic. Simple but loses type safety.
2. Tagged stack: Add runtime type tags. More Forth-like but adds overhead on GPU.
3. Separate float stack: Traditional Forth has a separate FP stack. Clean separation but doubles the plumbing.
4. Type-parameterized stack: Make stack element type configurable. Cleanest MLIR approach but bigger refactor.

Implementation notes

• Float arithmetic maps to arith.addf, arith.subf, arith.mulf, arith.divf
• Float load/store needs memref with f32/f64 element type, or bitcast at boundaries
• Float literals need parser support (detecting . or e in numbers)

Priority

Important — needed for any real-world matmul or ML kernel.