[ty] Compact indexed AST node storage (#25998)
## Summary
`IndexedModule` retains one lookup entry for every indexed AST node.
Prior to this change, each entry was a full `AnyRootNodeRef`, even
though resolving a `NodeIndex` only requires the node's address and root
AST kind.
An earlier version of this PR chose one packed, split, or wide
representation for the entire module. Corpus data showed that the split
layout was still used by 23.2% of modules, while the module-wide base
also meant that one distant allocation could widen the representation
for the entire AST. We explored a chunked replacement in #26047 and have
now folded that implementation back into this PR.
The final representation divides the index into fixed groups of 64
nodes. Each chunk chooses its own base address and the minimum entry
width needed for its nodes. Aligned relative offsets and five-bit
root-node kinds share a packed `u64` bitstream, while unusually large or
unaligned chunks use a local wide fallback. Deriving the chunk directly
from the node index keeps lookup constant-time, prevents unrelated AST
allocations from affecting one another, and removes the previous split
representation.
The review cycle also tightened the raw-pointer boundary.
`AnyRootNodeRef` now decomposes to `NonNull<()>`, its unsafe
reconstruction API documents the complete caller contract, and
`IndexedNodes` records the address-stability, provenance, liveness, and
node-kind invariants used during lookup. Collection retains typed
`AnyRootNodeRef` values until encoding, and the coverage round-trips
real indexed AST pointers rather than fabricated addresses.
The AST generator emits `RootNodeKind` and the raw-parts conversions
alongside `AnyRootNodeRef`, keeping the unflattened root-type mapping
synchronized as the AST evolves. This is intentionally separate from
`NodeKind`, which distinguishes concrete variants such as
`StmtFunctionDef` rather than the `Stmt` root type addressed by the
index. The traversal also stops indexing annotation expressions twice,
since `walk_annotation` already delegates to `visit_expr`.
The latest memory report for the final chunked implementation shows a
14.7–14.9% reduction in retained `parsed_module` memory across the
measured projects, reducing overall memory by 0.6–5.1%.
