Math function library

A library of LLVM IR function definition emitters for common math functions.

Exact mechanisms TBD, but currently:

• Rewrite via llvm::VecDescs
• Leave function calls and bodies intact until final opt pass for readability
• Attempts to be useful to GPU and both CPU pipelines
• Tracks accuracy with TPUs

Comparison with existing math libraries in XLA:

• libc, libm external function calls
  • Function call overhead
  • LLVM can't optimize across function boundaries
• cpu/codegen/polynomial_approximations
  • Rewrites inline
  • No optimizations after
• MLIR PolynomialApproximations.cpp
  • Uses MLIR RewritePatterns to implement the pass
  • Rewrites inline
  • Approximation accuracy goals independent of TPU