Split & remove extension_parallel

[swolchok]

As per plan in #8932, we want to be able to include thread_parallel.h to build
libraries that are capable of parallelization, but don't require it. So, we
move the header to ExecuTorch core and add a fallback implementation (with
tests!) of parallel_for that just does a regular for loop. Then, targets
that link extension_threadpool will get parallelization automagically.

This PR doesn't add any optionally-parallelized code; that will be in the next
PR.

[swolchok]

Stack from ghstack (oldest at bottom):

• Skip grain_size computation in parallel_for_each_reduce_* wrappers without threadpool #9144
• Add parallel_for_each_reduce_over_dim_list_output_index #9143
• Use parallel_for_each_reduce_over_dim_output_index for {map_,}reduce_over_dim ops #9141
• Extract parallel_for_each_reduce_over_dim_output_index from argmin parallelization PoC #9139
• use numel() when appropriate in get_reduced_dim_product #9142
• Proof-of-concept: parallelize argmin #9066
• Build optimized_portable_kernels if threadpool is enabled #8987
• parallelize optimized op_where using parallel_for #9059
• Use template instead of duplicating code in elementwise_util.h #9058
• Add BroadcastIndexesIterator::operator+ #9057
• Split delinearize_index out of broadcast_util #9056
• Make serial parallel_for "polyfill" iterate backwards in debug builds #9044
• -> Split & remove extension_parallel #8983

[swolchok]

ping reviewers @larryliu0820 @kirklandsign @lucylq @jackzhxng @iseeyuan @GregoryComer @kimishpatel