Unroll recursion in RuleCondition::matches This trades allocations due to Box::pin calls during recursion for an explicit stack and a tiny interpreter loop. (#79310)

### What?
Replace recursion with an explicit stack to reduce allocations.  Also, add tests!

This trims peak heap in a build of `vercel-site` by 100-300m, see https://vercel.slack.com/archives/C06PPGZ0FD3/p1747438469043789?thread_ts=1746823119.596949&cid=C06PPGZ0FD3


### Why?
This replaces a lot of `Box::pin` calls during evaluation, instead we can amortize the heap allocation cost by pushing items onto explicit stack. This is managed as a SmallVec so in many cases we should be able to avoid allocations.

### How?
Tediously introduce a small bytecode machine... sigh.  The main alternatives considered were:

* pulling all `awaits` up to the top level and using a normal recursive function.  This works, but would be limiting when we inevitably add another async dependency on the `Source::content`
* eliminating asynchrony from `Source::ident`.  This turned out to be controversial (see discussion in linear) and incredibly complex to implement so it is deferred.

Closes PACK-4560

Author: lukesandberg

Cargo.lock

@@ -54,6 +54,7 @@ rstest_reuse = "0.5.0"
 tokio = { workspace = true }
 turbo-tasks-malloc = { workspace = true, default-features = false }
 turbo-tasks-memory = { workspace = true }
+turbo-tasks-backend = { workspace = true }
 
 [build-dependencies]
 turbo-tasks-build = { workspace = true }