[autotuner] Introduce BenchmarkProvider abstraction for kernel benchmarking

[hinriksnaer]

sets up foundation for pytorch#1803

stacked on pytorch#1925

Motivation

Establishes a substitution point for alternative benchmarking strategies (e.g. cross-node precompilation, overlapped precompile+benchmark). Search algorithms call self.benchmark_provider.benchmark_function() instead of self.benchmark_function(). The provider owns the single-config benchmarking pipeline while _benchmark, benchmark_batch, and create_precompile_future remain on BaseSearch for follow-up migration. Once the changes are finalized, overall maintainability and effort required to introduce more fine grained stability measurements should be improved.

Summary

• Introduce BenchmarkProvider abstract class and LocalBenchmarkProvider implementation
• Move benchmark_function and its dependencies (baseline computation, accuracy validation, precompile management) from BaseSearch into the provider
• BaseSearch accepts benchmark_provider_cls parameter to allow alternative benchmarking strategies
• No behavioral changes, same code runs through the provider instead of directly on BaseSearch

Follow-up work

• Migrate _benchmark and create_precompile_future to the provider
• Migrate set_adaptive_compile_timeout as post_initial_benchmark
• Decouple accuracy validation from the benchmarking provider

Discussion

This is a step towards decoupling benchmarking strategies from search algorithms. The goal is to enable alternative benchmarking strategies without them being tied to individual search algorithm implementations, and to improve maintainability by giving benchmarking a clear boundary.

This PR moves benchmark_function and its direct dependencies into the provider as the smallest self-contained unit that establishes the abstraction without doing a comprehensive rewrite. The remaining methods on BaseSearch are marked with TODOs for incremental migration.

Given the scope of these changes, any feedback is very much appreciated.