ref(perfetto): Rework perfetto conversions

[Dav1dde]

No (or very minimal) functional changes intended.

• Better de-duplication, slightly more efficient through HashTable use for images, frames and stacks, instead of just frames.
• Slightly less allocations, again through HashTable referencing strings already allocated elsewhere.
• Introduction of builder types, used to build up our own sample format.
• Introduction of cache types, used just to cache data for a single sequence id
• InternTables is no longer also abused for caches.
• More modules to have proper type visibility, important for a follow-up to introduce better/stricter limits.
• Rework of most of the Rust code to actually make it Rusty not Codex Rust.
• Introduces a StackId and FrameId new-type wrapper instead of passing indices around, may prevent a bug at some point.
• Removes a bunch of useless extra intermediate types (like ResolvedFrame)
• Also happens to fix the precision for some timestamps, less precision loss (even if ms precision is enough).
• Caches are retained until there is a clear flag, instead of always clearing.

And yes there are some off-by-ones when it comes to the limits, they don't matter, if it messes with someone's OCD, I can fix them.

[loewenheim]

Looks very good, just some minor things.

[loewenheim]

Why is a 66-second chunk relevant in particular?

[Dav1dde]

Presumably due to the max profile length we support:

/// For continuous profiles, each chunk can be at most 1 minute.
/// In certain circumstances (e.g. high cpu load) the profiler
/// the profiler may be stopped slightly after 60, hence here we
/// give it a bit more room to handle such cases (66 instead of 60)
const MAX_PROFILE_CHUNK_DURATION: Duration = Duration::from_secs(66);