ZJIT: send_without_block_polymorphic happens too frequently

[k0kubun]

On lobsters, 19.4% of dynamic dispatch is caused by send_without_block_polymorphic.

We should:

1. Investigate the most frequent reasons why callsites become polymorphic. Try --zjit-trace-exits?
2. Address common polymorphic callsite patterns.
   • Support handling multiple receiver classes for a single callsite and then fallback to dynamic dispatch like YJIT does?

[tekknolagi]

I think at least a small part of this is because we might have a polymorphic cache of (class, shape, flags) list that is monomorphic in the class. So for the rewrites to SendWithoutBlockDirect, we only need to care about the cache being monomorphic in the first element, the class pointer. (For other rewrites, such as attr_accessor/attr_reader/attr_writer inlining, we need to look at the shape and flags too.) So there is some API missing such as is_monomorphic_class

[k0kubun]

I see. It seems like a nice thing to fix as a first step 🙂