Optimize hparams to work with data providers natively

[wchargin]

As of #3419, the hparams plugin uses the data provider APIs to read
data, but some of its internals still operate on the multiplexer data
structures (TensorEvent, TensorProto, etc.). These should be
migrated to just use the data provider APIs throughout, so that we don’t
have to convert back and forth. Callers that call read_scalars
repeatedly should also be migrated to call it just once with a
RunTagFilter representing the desired cross product, because this has
two orders of magnitude less processing overhead and (for networked data
providers) is one RPC instead of O(|metrics| × |sessions|) many.

[wchargin]

Partially addressed by #3438 (for prod code, not for test mocks).

[wchargin]

#3438 and #3449 finished this (plus #3478, I guess).