Reduce number of expression tree traversals in funcletization and caching

[roji]

Note: the following are some perf observations/ideas as a result of looking at the code - no profiling was done to ensure that this is a perf-critical area in the grand scheme of things. We should profile carefully before doing anything.

---

Following is a list of full tree traversal that we do for each query we execute (before compilation):

Parameter extraction:

• One pass with EvaluatableExpressionFindingExpressionVisitor to find evaluatable expressions NOTE: This was optimized with the funcletizer rewrite in Rewrite the funcletizer to support precompiled queries #33106
• One pass with ParameterExtractingExpressionVisitor to actually extract the parameters

Query caching:

• One pass to calculate the hash code of the query
• One pass per query in the cache with the same hash code, for Equals

So we're traversing the query tree at least four times for a cached query. We may be able to improve this by:

• Use a trie structure instead of the current dictionary (IMemoryCache), similar to trie-based string lookups.
• Merge parameter extraction with the trie traversal. As we're traversing the tree to find a cached entry, we also perform parameter extraction.

We should keep possible optimizations like this in mind when determining what is public and what isn't, to avoid breaking changes.

[roji]

Note: the traversals above can be avoided today via compiled queries, so this is a lower-priority area.

[roji]

Removal of the double visitation was done in #32698 - the new funcletizer works as a single visitors instead of as two.