[ty] Reduce size of member table

[MichaReiser]

Summary

This PR shrinks the size of the members table by ~28% (from 611MB to 438MB on a large project).

This is achieved by changing how we store MemberExpr.

Before

pub(crate) struct MemberExpr {
    symbol_name: Name,
	segments: SmallVec<[MemberSegment; 1]>,
}

pub(crate) enum MemberSegment {
    /// An attribute access, e.g. `.y` in `x.y`
    Attribute(Name),
    /// An integer-based index access, e.g. `[1]` in `x[1]`
    IntSubscript(ast::Int),
    /// A string-based index access, e.g. `["foo"]` in `x["foo"]`
    StringSubscript(String),
}

There are a few issues with this layout: MemberSegment itself is 32 bytes because it stores a String (or int) for each segment. This is an issue for multiple reasons:

1. We can only store a single segment inline
2. Even then, the small vec with only a single inline segment accounts for 40 bytes (32 bytes for a single segment + its metadata).
3. Each heap allocated segment also takes 32 bytes

This PR shrinks the size of MemberExpr and MemberSegment and reduces the cases where heap allocating the segments is necessary.

New

pub(crate) struct MemberExpr {
    /// The entire path as a single Name
    path: Name,
    /// Metadata for each segment (in forward order)
    segments: Segments,
}

The first change is that we change symbol_name to a path. The path stores the entire path of the member expression. For example, a.b[4] becomes ab4. This avoids the need for each segment to store its own Name (or Int)

The second important change is that Segments is now an enum over a small and a heap variant:

#[derive(Clone, Debug, PartialEq, Eq, get_size2::GetSize)]
enum Segments {
    /// Inline storage for up to 7 segments with 6-bit relative offsets (max 63 bytes per segment)
    Small(SmallSegments),
    /// Heap storage for expressions that don't fit inline
    Heap(Box<[SegmentInfo]>),
}

SmallSegments packs up to 7 segments into a single u64 (see docstring for more details).

SegmentInfo packs both an offset and a kind, so that a single segment only takes 4 bytes (it's a u32).

All these improvements together shrink MemberExpr from 64 bytes to 40 bytes

[github-actions]

mypy_primer results

No ecosystem changes detected ✅

Memory usage changes were detected when running on open source projects

trio (https://github.com/python-trio/trio)
- TOTAL MEMORY USAGE: ~167MB
+ TOTAL MEMORY USAGE: ~159MB

sphinx (https://github.com/sphinx-doc/sphinx)
-     memo fields = ~214MB
+     memo fields = ~204MB

[github-actions]

ruff-ecosystem results

Linter (stable)

✅ ecosystem check detected no linter changes.

Linter (preview)

✅ ecosystem check detected no linter changes.

Formatter (stable)

✅ ecosystem check detected no format changes.

Formatter (preview)

✅ ecosystem check detected no format changes.

[github-actions]

Diagnostic diff on typing conformance tests

No changes detected when running ty on typing conformance tests ✅

[carljm]

Nice!