Updating cache for INFORMATION_SCHEMA.INNODB_LOCKS et al is suboptimal

The function trx_i_s_common_fill_table() is filling and using a common cache for the following views:

• INFORMATION_SCHEMA.INNODB_TRX
• INFORMATION_SCHEMA.INNODB_LOCKS
• INFORMATION_SCHEMA.INNODB_LOCK_WAITS

This may invoke the function add_lock_to_cache() several times, once for each qualifying lock. There are a few problems with this:

• The lock is first being searched for, to avoid duplicates. (Why not do this as part of the insert? Could we use std::unordered_map::emplace() with a custom allocator?)
• The function fold_lock() may be invoked unnecessarily many times. There is also a similar problem in ha_storage_put_memlim(). It could be better to use my_crc32c(), like we did in MDEV-25905 for another hash table. In that way, the inefficient and collision-prone ut_fold_ulint_pair() and ut_fold_binary() would only be used for backward compatibility in the following:
  • databases that were originally created with innodb_checksum_algorithm=innodb; the default was changed already in MySQL 5.6 and MariaDB Server 10.0
  • the adaptive hash index functions dtuple_fold() and rec_fold(); changing that hash function might have an unpredictable performance impact