PS-10595 [9.7]: Reduce rw_lock_list contention during buffer pool initialization

[polchawa-percona]

https://perconadev.atlassian.net/browse/PS-10595

This is a contribution from Alexey Bychko abychko@gmail.com, with a
modification on top of the original patch: instead of inserting each chunk's
rw-locks into the global rw_lock_list one by one while holding
rw_lock_list_mutex, the per-chunk list is prepared outside the mutex and then
concatenated onto rw_lock_list in O(1) under the mutex.

During buffer pool initialization every buffer block rw-lock (block->lock
and, in debug builds, block->debug_latch) was registered in the global
rw_lock_list under rw_lock_list_mutex, one lock at a time. With large
buffer pools and parallel initialization this serializes heavily on that
single mutex.

Initialize block rw-locks without registering them in rw_lock_list, then
register a whole chunk's locks at once:

[+] rw_lock_init_only() initializes an rw_lock_t exactly like
rw_lock_create_func() but does not add it to rw_lock_list. A PFS-aware
wrapper pfs_rw_lock_init_only_func() and the rw_lock_init_only_inst()
macro preserve placement-new, pfs_psi and (debug) latch id handling.
[+] ut_list_concatenate() / UT_LIST_CONCATENATE() splice one UT_LIST onto
another in O(1).
[+] rw_lock_list_register_bulk() builds nothing under the lock: the caller
prepares the per-chunk list outside rw_lock_list_mutex, and this function
only splices it onto rw_lock_list in O(1) under a single mutex section.
[*] buf_block_init() now uses rw_lock_init_only_inst(); buf_pool_register_chunk()
collects the chunk's block rw-locks into a local list and registers them
in bulk.

rw_lock_create_func() keeps its existing signature and behavior, so all
other callers and diagnostics based on rw_lock_list are unaffected; the
complete rw_lock_list is available once initialization finishes.