fix match_modules_set to work with MoE

[HDCharles]

Summary:

match_modules_set isn't currently as useful as it could be because it lacks the ability to match multiple results for each target like in the case of a moe model where you have multiple experts.

all need to be matched when doing something like AWQ or spin quant.

```python3
targets = [
    "post_attention_layernorm", 
    "up_proj", 
    "down_proj"
]
match_modules_set(model, targets) == (
    [
        [layers.0.post_attention_layernorm],
        [
            `layers.0.mlp.experts.0.up_proj`,
            `layers.0.mlp.experts.1.up_proj`,
            ...

        ],
        [
            `layers.0.mlp.experts.0.down_proj`,
            `layers.0.mlp.experts.1.down_proj`,
            ...
        ]
    ], # <- first yield
    [<same but for layers.1>] # <- second yield
    ...
)
```

In order to make is so this can still work for matching simple cases like qkv and moe cases, we use the following approach:

Algorithm

1. match modules until we have at least 1 match per target
2. when we have 1 match per target, our set is 'full' and we calculate
   the common parent context
3. continue matching and for each match, check if parent context would change given the
   new match
4. if we find a match that changes the parent context, this is the first
   element of the next set. yield the existing matched set and then
   reset, using the current match as the first element of the new set.

Requirements

A) one element from each target is enough to define the parent context of a set, i.e. adding more elements from the same set wouldn't change the parent context once its initially set.
B) each set would have a different parent context, or else we'll just put everything into one huge set.

Other

To facilitate this algorithm i also added get_lowest_common_ancestor_name which basically copies a similar function in llm-compressor though significantly simpler.

[kylesayrs]

Consider adding a test for the MoE case. Also remember to change the existing use case after landing.

Really impressive work, thanks

[fynnsu]

Nice algorithm! Left a couple comments below

[kylesayrs]

Neat and cool and neat and cool

[fynnsu]

Looks good!

[fynnsu]

Approved but with an optional suggestion

[HDCharles]

Approved but with an optional suggestion

I would have fixed it if it wouldn't get rid of the review!