Description: Simplify the range of parallel loops implemented in the thread pool by removing the variants based on SchedulingParams. This leaves the variant that takes a simple double to express costs, and the variant that takes a TensorOpCost struct.

Motivation and Context
The parallel loop variants taking a SchedulingParams were essentially unused, but supporting them introduced complexity in writing and testing other changes to the thread pool implementation. The single use is in the gelu.cc microbenchmark which uses SchedulingParams to force a fixed 4096-size chunk for work distribution. I updated the microbenchmark to do this chunking explicitly, following the code in onnxruntime/contrib_ops/cpu/bert/bias_gelu.cc