[FIRRTL][Dedup] Alter dedup group handling, avoid exponential behavior.

[dtzSiFive]

Transmute these annotations to simple (temporary) attributes on modules.
This causes the desired behavior for hashing and since this must match, when deduplicating no work is needed.

Presently, as annotations, each module dedup'd into the group walks all annotations and "adds context" (makes them non-local) and interns a new annotation array with those annotations. At the end of the pass, all dedupGroup annotations are removed.

This causes a lot of unnecessary symbols, hierpaths, annotations, and arrays of annotations as well as some quadratic behavior.

DedupGroup annotations are commonly on every single module, so for those designs large dedup groups scaled poorly.

Instead, drop these annotations immediately and add as a simple named attribute indicating the group.

Drop the (temporary) dedup group attributes at end of pass, as they are just an in-IR way to track this per-module during the pass.

Fixes #6979.

---

Benchmarking (n=10, median) on some quick synthetic examples put together, FWIW (top has instances of the many identical empty modules):

Name	Modules	Instances	real before (s)	real after (s)	Before/After	maxrss before (kb)	maxrss after (kb)	Before/After
Groups	3001	3000	1.035	0.311	3.33	124,096	72,944	1.70
Groups_3x	3001	9000	2.237	0.361	6.20	216,696	77,456	2.80
Many	10001	30000	33.349	3.461	9.64	1,802,784	475,366	3.79

Significantly improves performance where there are many instances of deduplicated modules and there are dedup group annotations.

[youngar]

Just to connect the dots, this is similar to what we do to prevent memories in the testharness and (i think) prefix-module-groups from deduping with each other: #2968

edit: looks like the groupId stuff was replaced with string prefixs: #4952

[youngar]

As we talked about offline, one idea for the future is to change the annotation handler to attach this attribute (in LowerAnnotations) instead of doing it here. Change looks great to me, thanks!