docs: Enhance mixed-precision documentation and remove temporary scripts

Enhanced ConfigureMixedPrecision() documentation in both PredictionModelBuilder.cs
and IPredictionModelBuilder.cs to clearly explain technical constraints:

1. Type Constraint: float only
   - Mixed-precision converts between FP32 (float) and FP16 (Half)
   - Cannot use double, decimal, or integer types

2. Gradient-Based Optimizers Only
   - Requires gradient computation for loss scaling, master weights, and gradient accumulation
   - Does NOT work with non-gradient methods (genetic algorithms, random search, Bayesian optimization)

3. Neural Networks (Recommended)
   - Best suited for networks with large parameter counts
   - Requires GPU with Tensor Core support for 2-3x speedup
   - Provides 50% memory reduction for massive models

Also removed temporary development scripts from scripts/ directory:
- add-half-conditional*.py (conditional compilation helpers)
- add-half-ifdef.sh (development utility)
- check-encoding.sh (encoding validation)
- fix-encoding.py (encoding repair)
- fix-half-conditionals.py (development utility)

These were accidentally committed during Half type conditional compilation work.
Kept launch-distributed-training scripts as they are part of the public API.

Author: claude