Handle refs/convert/parquet and PR revision correctly in hffs

[Wauplin]

Solves #1710 by processing refs/convert/parquet and refs/pr/(\d)+ as special revisions when resolving a path. Also make sure that the unresolve counterpart works the same. Also added some tests for it.

Use case from @Hakimovich99 now works properly without having to url-encode the revision. This time I tested with jamescalam/llama-2-arxiv-papers-chunked dataset which doesn't have parquet files on main revision:

>>> from dask import dataframe as dd
>>> dd.read_parquet("hf://datasets/jamescalam/llama-2-arxiv-papers-chunked@refs/convert/parquet")
# Dask DataFrame Structure:
#                  doi chunk-id   chunk      id   title summary  source authors categories comment journal_ref primary_category published updated references
# npartitions=1                                                                                                                                              
#               object   object  object  object  object  object  object  object     object  object      object           object    object  object     object
#                  ...      ...     ...     ...     ...     ...     ...     ...        ...     ...         ...              ...       ...     ...        ...
# Dask Name: read-parquet, 1 graph layer

>>> dd.read_parquet("hf://datasets/jamescalam/llama-2-arxiv-papers-chunked@refs/convert/parquet/default/train/0000.parquet")
# Dask DataFrame Structure:
#                   doi chunk-id   chunk      id   title summary  source authors categories comment journal_ref primary_category published updated references
# npartitions=1                                                                                                                                              
#                object   object  object  object  object  object  object  object     object  object      object           object    object  object     object
#                   ...      ...     ...     ...     ...     ...     ...     ...        ...     ...         ...              ...       ...     ...        ...
# Dask Name: read-parquet, 1 graph layer

This should be a small QOL improvements for datasets users while not degrading the experience for others.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[lhoestq]

Btw forgot to mention this in the other issue but using the ~parquet alias works:

dd.read_parquet("hf://datasets/jamescalam/llama-2-arxiv-papers-chunked@~parquet")

[Wauplin]

Btw forgot to mention this in the other issue but using the ~parquet alias works:

Aaaah, right. Completely forgot as well. I think that's why I thought we already talked about simplifying the revision (as mentioned in #1707 (comment)). Anyway, I think handling parquets + PR refs differently is also good in addition to of the ~parquet alias.

[lhoestq]

Looks good to me !

[Wauplin]

Thanks for the review! Failing tests are unrelated so I'll merge now :)