Fix unimplemented functions mapping and CUDA TF32 precision setup

[janpfeifer]

This PR introduces two main enhancements/fixes:

1. Unimplemented Functions Integration:

   • Embeds notimplemented.Function into xla.Function.
   • Adds a newFunction helper to set up standard XLA error handling, raising an informative message ("{op} not implemented for XLA backend") when an unimplemented operation is called on the function.
   • Refactors b.Main(), b.NewFunction(), and f.Closure() to use newFunction.

2. CUDA TF32 Precision Handling:

   • Updates DotGeneral precision settings for CUDA:
     • Explicitly requests DotGeneralPrecisionDefault when DotGeneralUseTF32 is enabled and accumulation dtype is F32 (retaining TF32 performance).
     • Otherwise, falls back to DotGeneralPrecisionHighest for maximum accuracy.
   • Sets DotGeneralUseTF32 to enable by default on CUDA backends (isPluginType(pluginName, "cuda")).