Dataset subsets

[voegtlel]

Fixes #13

Implements the sample range like this:

Example from tests:

ds1: Values [0, 55]
ds2: Values [100, 155]
ds3: Values [200, 255]

metadataset_ratio2.yaml:

__module__: megatron.energon
__class__: MetadatasetV2
splits:
  train:
    # take [10, 30] from ds1, [20, 40] from ds2 and then only [20%, 80%]
    # I.e. sample range: [14, 26], 2 * [124, 136]
    subset: {range: [20%, 80%]}
    blend_epochized:
      - path: ds1
        subset: {range: [10, 30]}
      - repetitions: 2
        subset: {range: [20, 40]}
        path: ds2

__module__: megatron.energon
__class__: MetadatasetV2
splits:
  train:
    subset_ratio: [0%, 50%]
    blend_epochized:
      - path: ds3
        # take [30, 50] from ds3, then first 50%, resulting in samples [230, 240]
        subset: {range: [30, 50]}
      - repetitions: 2
        # Inner sample range: [14, 26], 2 * [124, 136], total=12*3=36
        # Applying subset ratio 25%-75%: [17, 23], 2*[127, 133], total=3*6=18
        # Applying outer 50%: [17, 20], 2*[127, 130], total=3*3=9
        # Applying repetition: 2*[17, 20], 4*[127, 130], total=2*9=18
        subset: {range: [25%, 75%]}
        path: metadataset_ratio2.yaml

Recursively applies the subset (innermost first). subset with absolute values can only be applied to a "leaf" dataset (i.e. not a recursive metadataset).

This is

Creating as draft up for discussion.