Fix shard computation in NoShuffleBeamWriter with overlapping split names

[lgeiger]

Currently NoShuffleBeamWriter would compute the wrong number of shards for cases where splits have overlapping names.

E.g. given a dataset with two splits foo and foo_bar the generated tfrecord files would look something like this:

builder-foo.tfrecord-00000-of-00003
builder-foo.tfrecord-00001-of-00003
builder-foo.tfrecord-00002-of-00003
builder-foo_bar.tfrecord-00000-of-00002
builder-foo_bar.tfrecord-00001-of-00002

The current regex for the foo split would be builder-foo* which also matches the files of the foo_bar split. This results in the number of shards for foo to also include shards from the other split. This fixes it by changing the regex to builder-foo.*.

[lgeiger]

@tomvdw Do you mind having a look at this?