perf(serde_v8): drop need for RC in serializer

[AaronO]

Which substantially improves the performance of to_v8() by removing RC's overhead (extra allocs, drops, checks, etc...)

This breaks through the symbolic barrier of 200ns/op overhead (on my machine) ! 🚀 🚀 🚀
(The last symbolic milestone before going sub 100ns/op, which will be very hard to cross ...)

Benches

Note that the before and after of these benches integrates changes from #9983.

Before:
test bench_op_async   ... bench:     716,780 ns/iter (+/- 28,508)
test bench_op_nop     ... bench:     213,206 ns/iter (+/- 5,618)
test bench_op_pi_bin  ... bench:     243,481 ns/iter (+/- 6,120)
test bench_op_pi_json ... bench:     227,972 ns/iter (+/- 5,995)

After:
test bench_op_async   ... bench:     647,272 ns/iter (+/- 57,528)
test bench_op_nop     ... bench:     157,173 ns/iter (+/- 2,096)
test bench_op_pi_bin  ... bench:     179,239 ns/iter (+/- 4,354)
test bench_op_pi_json ... bench:     175,075 ns/iter (+/- 1,814)

[bartlomieju]

LGTM