PS-11428 get_rnd_index is limited by TSC resolution in newer CPU architectures

[marcinbabij]

[1/2]

Problem:
With innodb_sync_array_size > 1, sync_array_get_and_reserve_cell
can hit ut_a(*cell != nullptr) and abort even though another
wait-array instance still has a free cell. A cell is expected
whenever any instance has space.

Root cause:
The retry loop calls sync_array_get() independently on every
iteration, up to sync_array_size times. That is sampling with
replacement: the same full instance can be chosen on every
try, so a free cell elsewhere is never visited. On 1024
instances with one free, P(miss) is (1023/1024)^1024 ~ 1/e
per call.

Fix:
Walk every instance once from a random start so a full
instance cannot hide a free cell in another instance.

[2/2]

Problem:
get_rnd_index() feeds sync wait-array start, ib_counter_t slots,
and sharded rw-lock S-latch. On CPUs whose RDTSC increments by
32, index % 32 is always 0, so those users hit one bucket.

Root cause:
my_timer_cycles() is used as the index with no mixing. A
constant TSC stride leaves low bits stuck.

Fix:
XOR the sample with itself shifted by 6. One shift and one
XOR after RDTSC. Covers strides up to 64 and keeps
increment-1 entropy.

Test plan

• sync0arr-t
• ut0rnd-t --gtest_filter='ut0counter.*'