[Feature] Support KLD metric and support evaluation for probabilistic models

[LeoXing1996]

Metric Design

Different from gan metrics, probabilistic metrics:

1. calculated with reconstruction tasks, and can
2. calculated by a group of parameters but not fake and real images

Therefore, we design a list called probabilistic_metric_name to contain probabilistic metrics in single_gpu_online_evaluation . When evaluation with those metric separately.

When evaluating those probabilistic metrics, we use set forward mode as reconstruction.
We default that all probabilistic model all support this mode. Unlike forward_test which starts with random noise, mode=reconstruction performs a reconstruction behavior with given data and returns a dict containing desired probabilistic parameters.

We also slightly modify the batch truncation operation to make the Metric.feed support dict input.

Some further design about mode=reconstruction

Although we have not implemented any code, we find a specific function to release reconstruction operation is critical for probabilistic models (e.g., DDPM).
In train_step, reconstruction, loss calculation, and update operation are performed altogether.
For forward_test, the interface is fixed and called by sample_from_noise to perform a random generation process.
Therefore, we need a function to implement a separate reconstruction process and return all the intermedia probabilistic parameters. And this function can also be called by train_step.

[codecov-commenter]

Codecov Report

Merging #108 (1146a8f) into master (f8d6f2e) will decrease coverage by 0.78%.
The diff coverage is 77.25%.

❗ Current head 1146a8f differs from pull request most recent head 736a71b. Consider uploading reports for the commit 736a71b to get more accurate results

@@            Coverage Diff             @@
##           master     #108      +/-   ##
==========================================
- Coverage   76.06%   75.28%   -0.79%     
==========================================
  Files         118      121       +3     
  Lines        8089     8188      +99     
  Branches     1519     1561      +42     
==========================================
+ Hits         6153     6164      +11     
- Misses       1546     1609      +63     
- Partials      390      415      +25     

Flag	Coverage Δ
unittests	75.28% <77.25%> (-0.79%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
mmgen/core/evaluation/evaluation.py	7.30% <0.00%> (-0.98%)	⬇️
...rchitectures/sngan_proj/generator_discriminator.py	88.10% <ø> (ø)
mmgen/models/gans/__init__.py	100.00% <ø> (ø)
mmgen/models/gans/base_gan.py	68.13% <ø> (ø)
mmgen/models/losses/utils.py	83.33% <0.00%> (-5.96%)	⬇️
mmgen/utils/collect_env.py	20.00% <0.00%> (-1.16%)	⬇️
mmgen/datasets/unpaired_image_dataset.py	83.33% <72.72%> (-1.29%)	⬇️
...odels/translation_models/static_translation_gan.py	74.57% <74.57%> (ø)
mmgen/core/evaluation/metrics.py	75.36% <76.92%> (-0.04%)	⬇️
mmgen/apis/inference.py	47.31% <81.81%> (-1.11%)	⬇️
... and 18 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f8d6f2e...736a71b. Read the comment docs.