EXPERIMENTAL: feat(ops): String fast ops, fast op_base64_decode

[aapoalas]

This PR introduces String support to fast ops, and trials a method of making ZeroCopyBuf-returning ops fast-compliant by having them return a ResourceId and (through an out-buffer) a buffer size. JS side will then create a buffer of that size and call a finisher op that writes the Vec to the passed-in buffer and drops the Vec.

The fast-compliant ZeroCopyBuf returning functions are a pretty easy idea and shouldn't really cause any issues. The experimental nature of this PR is in the String support for fast ops.

First of all, the support required adding a v8::Isolate pointer to the OpState which is already really dirty. Additionally, while @littledivy has verified that it is possible to get the string content, it may not work in all cases. @AaronO raised the possibility that getting the string content for string ropes might cause V8 to do string combining, meaning that suddenly the fast op would be proccing heap allocations and that might cause V8 to panic.

Locally tests pass so nothing is breaking directly, but tests likely never use roped strings so this is no guarantee. Additionally, op_base64_decode is apparently not used at all, anywhere.

[bartlomieju]

@littledivy can we land this PR?

[aapoalas]

No

[aapoalas]

I'll close this for now, at least.

The branch did show fast calls using Strings, but apparently not during benchmarking. I could not see any improvements in performance.

The most important contribution in this branch is probably the work to eliminate a DoubleBorrow error from ops/lib.rs, but I've already mostly forgotten which change was actually the relevant part fixing the issue.