[Core][DO NOT MERGE] Improve OpenMP loop scheduling and chunk partitioning configurability in ParallelUtilities

[loumalouomega]

📝 Description

This PR updates shared-memory loop execution in ParallelUtilities to improve load balancing and runtime tunability, while keeping dynamic scheduling as the default fastest option from local benchmarking.

This is kind of a retake of #12923

NOTE: After discussing with @pooyan-dadvand we need further changes, and I may retake this later.

🚀Benchmarking

Intel® Core™ Ultra 9 Processor 285HX in Windows 11

Performance improvement for large sets, for small sets equal or slightly worse

Intel® Core™ Ultra 9 Processor 285HX in Ubuntu 24.04 via WSL

Consistent x8+ performance

🔀 What changed

• Added ChunkPartitioningScheme with two modes:
  • DIVIDE_BY_NUMBER_OF_CHUNKS
  • DIVIDE_BY_CHUNK_SIZE
• Extended BlockPartition and IndexPartition templates to accept the partitioning scheme as a template parameter.
• Updated partition constructors to accept a generic N (interpreted as chunk count or chunk size depending on scheme).
• Added global tuning knobs:
  • ParallelUtilitiesMaxChunkSize (default 1024)
  • ParallelUtilitiesMaxNumberOfChunks (default Globals::MaxAllowedThreads, then adapted at init)
• Added env-based runtime configuration in InitializeNumberOfThreads():
  • KRATOS_PARALLEL_MAX_CHUNK_SIZE
  • KRATOS_PARALLEL_MAX_CHUNKS
• If KRATOS_PARALLEL_MAX_CHUNKS is not set, default is adjusted to min(4 * num_threads, ParallelUtilitiesMaxNumberOfChunks).
• Set OpenMP loop scheduling to schedule(dynamic) in all relevant loops in parallel_utilities.h.

📒Notes

• schedule(dynamic) is intentionally kept hardcoded for now based on benchmark results; switching to schedule(runtime) can be done later if runtime policy control is preferred.

😅 TODO

• Benchmark in different OS
• Benchmark in different CPU

🆕 Changelog

• Add environment variable support for maximum chunks in ParallelUtilities and consider dynamic schedulling
• Update default maximum chunks in ParallelUtilities to use Globals::MaxAllowedThreads)
• Set default value for maximum chunks in InitializeNumberOfThreads based on thread count
• Introduce chunk partitioning scheme and environment variable support for parallel utilities

[loumalouomega]

Maybe this conflicts with C++ implementation