Use single allocation to store origin and QueryRevisionExtra

[MichaReiser]

With #1101, the layout for a memo with extra data is:

QueryRevisions
+--------------------------------------+
| changed_at                           |
| durability / flags                   |
|                                      |
| OriginAndExtra (13 bytes)            |
| +----------+----------+------------+ |
| | tag      | payload  | metadata   | |
| | Indirect | pointer  |     0      | |
| +----------+----|-----+------------+ |
+-----------------|--------------------+
                  |
                  v
        Box<IndirectOriginAndExtra>
        +----------------------------------+
        | origin: OriginAndExtra           |
        | +----------+----------+--------+ |
        | | Direct   | pointer -+ | count| |
        | +----------+----------|-+------+ |
        |                       |          |
        | extra (24 bytes)      |          |
        | +------------------+  |          |
        | | tracked IDs      |  |          |
        | | cycle heads      |  |          |
        | | iteration/flags  |  |          |
        | +------------------+  |          |
        +-----------------------|----------+
                                |
                                v
                           Box<[E]>
                           +-----+-----+-----+-----+
                           | E 0 | E 1 | E 2 | ... |
                           +-----+-----+-----+-----+

Allocations: 2
  1. IndirectOriginAndExtra, approximately 40 bytes
  2. Edge slice, sizeof(E) * edge_count

Note how this requires two boxes. One to IndirectOriginAndExtra and another for the query edges.

This PR avoids the two boxes by using a custom DST that stores the QueryRevisionExtra as Header and the origin edges as slice, both using the same Box

QueryRevisions
+--------------------------------------+
| changed_at                           |
| durability / flags                   |
|                                      |
| OriginAndExtra (13 bytes)            |
| +------------+----------+----------+ |
| | tag        | payload  | metadata | |
| | With extra | pointer  | count    | |
| | Packed/Wide|    |     |          | |
| +------------+----|-----+----------+ |
+-------------------|------------------+
                    |
                    v
       SliceWithHeader<QueryRevisionsExtraInner, E>
       +--------------------+-----+-----+-----+-----+
       | extra (24 bytes)   | E 0 | E 1 | E 2 | ... |
       |                    |     |     |     |     |
       | tracked IDs        |     inline edges      |
       | cycle heads        |                       |
       | iteration/flags    |                       |
       +--------------------+-----------------------+

The memory gains on ty are marginal, but walltime of cycle heavy projects improve by 10% (20% fewer allocations)!
astral-sh/ruff#25774

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	072d3ae
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/6a27051ac8836f00093d73a8

[netlify]

✅ Deploy Preview for salsa-rs canceled.

Name	Link
🔨 Latest commit	bf11195
🔍 Latest deploy log	https://app.netlify.com/projects/salsa-rs/deploys/6a2915513b582500081fd9c8

[codspeed-hq]

Merging this PR will improve performance by 5.79%

⚡ 2 improved benchmarks
✅ 24 untouched benchmarks

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Memory	converge_diverge	9.1 KB	8.6 KB	+6.28%
⚡	Simulation	converge_diverge	170.5 µs	161.9 µs	+5.3%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing MichaReiser:micha/co-allocate-query-revision-extras (bf11195) with master (f6325bb)}

[carljm]

Looks reasonable to me! Neat optimization.