Fix performance hang and OOM on conditional tuple appends

[QEDady]

Summary

This PR fixes a performance hang and Out-Of-Memory (OOM) error that occurs when type checking code with many conditional tuple appends (e.g. at control-flow joins).

Problem

At each control-flow join, Pyrefly computes the union of incoming types. For $N$ conditional tuple appends of distinct class types, this leads to an exponential combination of up to $2^N$ concrete tuple variants in the union. Pyrefly's simplifier stalls attempting to sort, deduplicate, and intersect this list, leading to $O(2^N)$ CPU time and OOM.

Solution

We introduce collapse_wide_tuple_unions during type simplification. If a union list contains more than a threshold of 8 tuple variants, they are collapsed/widened into a single unbounded tuple variant:
tuple[E, ...]
where E is the union of all element types recursively extracted from the constituent tuples (concrete, unbounded, or unpacked). This soundly bounds type complexity at control-flow joins to linear $O(N)$ complexity.

Fixes #3743

Test Plan

1. Rust Unit Tests: Added test_collapse_wide_tuple_unions_corner_cases in simplify.rs to verify correct widening of concrete, unbounded, nested, unpacked, TypeVarTuple, and invalid fallback tuple combinations.
2. Integration Regression Test: Added a 30-branch conditional tuple append integration test to test/basic.md using distinct class instances to ensure the widening threshold is properly triggered.
3. Execution: Ran the tests

   python3 test.py 

4. Performance Verification: Checked the 30-branch reproduction script on the local build. The run, which previously hung/timed out (exceeding 60 seconds at 100% CPU), now compiles successfully in 0.18 seconds with flat memory footprint (~360MB).

[yangdanny97]

Nice, thanks

[meta-codesync]

@yangdanny97 has imported this pull request. If you are a Meta employee, you can view this in D108143939.

[yangdanny97]

FYI I'm gonna bump the limit to 64 here

[yangdanny97]

or rather, 256... pytorch has an un-annotated list of ~100 2-tuples :P

[samwgoldman]

Review automatically exported from Phabricator review in Meta.

[meta-codesync]

@yangdanny97 merged this pull request in 1c5d158.