Measure the impact of prejitting/tiering on casting helpers.

[VSadov]

In local microbenchmarks it was observed that prejitting casting helpers has some impact on throughput.
We should investigate further whether the effect is real and not just some noise, and consider if we can improve anything.

[VSadov]

The regression in #37803 captures the worst cases scenario.
It is the simplest of the casts (casting a class with one implemented interface) and we have observed a statistical penalty up to 29%.

The reason why it is noticeable is because operation is itself very fast and that we measure a loop that performs 100000 casts, so we can see impact of tiering indirection.

In absolute terms the impact is very small and unlikely to become an issue in real world scenarios. It is possible to remove the indirection cost by making these helpers jitted, but it seems to be reasonable trade for the start up latency.