feat: add float math intrinsics

[tetsuo-cpp]

Summary

• Add 7 float math intrinsic words: FEXP, FSQRT, FLOG, FABS, FNEG (unary) and FMAX, FMIN (binary)
• Lower unary ops via math.exp/math.sqrt/math.log/math.absf/arith.negf with i64↔f64 bitcast pattern
• Lower binary ops via arith.maximumf/arith.minimumf reusing existing BinaryArithOpConversion template
• Add convert-math-to-llvm pass in pipeline for GPU lowering

Test plan

• Translation test: all 7 words parse to correct forth.* ops
• Conversion test: unary/binary ops produce expected bitcast + math/arith patterns
• Pipeline test: end-to-end lowering to gpu.binary
• All 105 tests pass

[tetsuo-cpp]

Code Review

Clean, well-structured PR that follows established patterns closely. LGTM overall.

Strengths

• Good template factoring — UnaryFloatOpConversion mirrors the existing BinaryArithOpConversion nicely
• MaxF/MinF correctly reuse the binary template instead of duplicating code
• Test coverage across all three layers (translation, conversion, pipeline)
• Pipeline ordering is correct: math-to-llvm after gpu-to-nvvm but before nvvm-to-llvm

One bug: duplicate stage number

Passes.cpp:53 — comment says "Stage 8" for gpu.binary compilation, but the preceding reconcile-unrealized-casts is also labeled "Stage 8". Should be "Stage 9".

Nit

arith::MaximumFOp / arith::MinimumFOp follow IEEE 754-2019 semantics (NaN-propagating). This is probably the right default for GPU math, but worth noting — if FMAX NaN 1.0 should return 1.0 (like CUDA's fmax), you'd want arith::MaxNumFOp instead. The current choice matches standard IEEE behavior though, so fine as-is.