Introduce randomized testing for queries, fix the revealed bugs

[maxbrunsfeld]

Randomized testing is already used in Tree-sitter's test suite for verify that incremental parsing behaves correctly. But previously, Tree-sitter's query engine was only tested via hand-written examples. This PR adds some limited randomized testing of the query engine.

Strategy

The new tests use the following procedure:

1. Parse one example file.
2. Pick a random node in the resulting syntax tree.
3. Generate a random query that matches that syntax node.
4. Parse a second example file.
5. Compute the "expected matches" for the random query and this second syntax tree, using a brute-force, backtracking approach.
6. Generate a list of "actual matches" by running the query itself on this second tree.
7. Check that the actual matches are the same as the expected matches.

Findings

This test found a number of patterns which gave false "impossible pattern" errors when constructing queries, due to bugs in the query analysis. They also surfaced a number of places where we were unnecessarily splitting match states, which ultimately resulted in incorrectly-ordered results.

Limitations

The test that's checked-in has some restrictions:

1. It only uses the Rust grammar.
2. It only draws from one particular example file
3. It does not generate patterns with optional nodes, repeated nodes, anchors, or alternatives
4. A new random seed is not selected every time you run the test suite: it always run the same 100 random examples. To explore more seeds, you have to change the test code.

Next Steps

After this, it'll be easy to gradually enhance the randomized tests to exercise more languages, more example files, and more of the query syntax. We should also test iterating through captures, not just matches.