[TTS] Infrastructure for parallelization of evaluation (scoring)

[rfejgin]

This PR makes MagpieTTS's evaluation a little faster and much more parallelizable.

Changes:

• Made the ASR step batched, rather than batch-size-1. We first collect all audios that need ASR and run ASR on them in a batched manner before the main per-sample evaluation loop. This substantially speeds up the ASR part of evaluation.
• Infrastructure towards multi-GPU evaluation (scoring). That is something we will do (and have prototyped) with NeMo Skills later on. To enable that, evaluation was broken down into two steps: a first step where evaluation of each utterance is independent of other utternaces, and a second step that focuses on parts that require with global state.
• During refactoring I also removed the constant STANDARD_METRIC_KEYS since the set of metrics on which to compute CIs can be inferred from the metrics themselves, which should be easier to maintain.

So evaluate() has been broken into:

1. evaluate_dir(): computes metrics for all audios in a given directory+manifest and outputs per-file metrics.
2. compute_global_metrics(): takes per-file metrics collected (1) and computes global metrics from those. This mostly amounts to computing averages. But it also includes computing FCD computation, since that is not something that can easily be broken down into directory-wise chunks due to the statefulness of the FCD metric.
3. evaluate(): wrapper that chains (1) and (2) for easy use in NeMo. NeMo skills would call (1) and (2) directly.

Running on a single GPU (local machine), these changes yield ~20% speedup of evaluation (more for larger sets, less for small ones, due to overhead of loading models). The benefit is much larger when evaluating on multiple GPUs in parallel, which we have prototyped in NeMo Skills (and will merge later on).

@vmendelev : adding you just as FYI. I think your existing Skills integration should work as-is after this PR, since the evaluate() API hasn't changed. Later, we can break down how Skills does Magpie scoring into the parallelizable part (evaluate_dir()) and the aggregation step (compute_global_metrics()) – I experimented with that and it seemed to work well.

[rfejgin]

This check no longer makes sense since the metric names are now inferred from the metrics themselves.

[rfejgin]

Removed for better maintainability - we can infer these names from the metrics themselves.

[rfejgin]

As part of making evaluation more efficient, we want to ensure UTMOS models don't get re-downloaded each time.

[blisc]

Hmm, guess we always hard-coded this, but should we remove this hardcode?

[rfejgin]

Done in latest commit. It's part of EvaluationConfig now, still defaulting to "cuda".