fix(data): resolve data_type aliases and 'auto' to dispatchable dataset ids

[SuperMarioYL]

Problem

DataLoaderFactory.create_data_loader documents data_type as accepting
"text", "multimodal" or "auto" (and "auto" is the parameter default),
but those values never work:

• Auto-detection assigns data_type = "text" or "multimodal", yet the
  dataset dispatch only matches the class-name keys ("TextDataset",
  "MultiModalDataset", "Text2ImageDataset", "OmniDataset",
  "AudioDataset").
• So data_type="text", data_type="multimodal", and data_type="auto"
  all fall through to the final else and raise
  ValueError: Unsupported data type: text.

Any caller that uses the documented short names — or simply relies on the
default data_type="auto" — hits this. Only the full class-name strings
supplied by the shipped YAML configs (name: TextDataset) happen to work.

Fix

Extract a small, pure _resolve_data_type(data_type, data_source) helper that:

1. performs the existing "auto" detection, then
2. normalizes the documented short aliases (text, multimodal, text2image,
   omni, audio) to their canonical dataset id via a DATA_TYPE_ALIASES
   map, and
3. validates the result against SUPPORTED_DATA_TYPES, raising a descriptive
   ValueError for anything unrecognized.

Class-name values supplied by existing configs pass through unchanged, so this
is fully backward compatible. The docstring is updated to list the values that
are actually accepted.

Tests

tests/test_dataloader.py — CPU-only (no GPU, no model weights; the heavy
torch/transformers and sibling-dataset imports are stubbed so the pure
resolution logic can be exercised in isolation, matching the style of the
existing tests/test_config_parser.py). 16 cases cover:

• "auto" resolving to a dispatchable id for .json/.parquet/dir/dict sources,
• every documented short alias,
• every canonical class name passing through unchanged (regression guard),
• the unknown-type ValueError.

Red on main (the documented values are not dispatchable), green on this branch.
black --line-length=99, isort --profile=black, and flake8 all pass.

Scope

+54 / -9 in angelslim/data/dataloader.py plus the new test file. No change to
the success path of configs that already pass class names.