Allow broadcasting across dimensions in eval mode; always require X to be at least 2d

[esantorella]

Summary:
Context:

A discussion on allowable shapes for transforms concluded:

• We should not allow for broadcasting across the -1 dimension, so the first check in _check_shape should always happen.
• The shapes always need to be broadcastable, so the torch.broadcast_shapes check in _check_shape should always happen.
• We want to allow for broadcasting across the batch dimension in eval model, so the check that X has dimension of at least len(batch_shape) + 2 should only happen in training mode.
• For clarity, we should disallow 1d X, even if broadcastable. BoTorch tends to be strict about requiring explicit dimensions, e.g. GPyTorchModel._validate_tensor_args, and that's a good thing because confusion about tensor dimensions causes a lot of pain.

This diff:

• Only checks that X has number of dimensions equal to 2 + the number of batch dimensions in training mode.
• Disallows <2d X.

Differential Revision: D62404492

[facebook-github-bot]

This pull request was exported from Phabricator. Differential Revision: D62404492

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.98%. Comparing base (33e11f4) to head (1618797).
Report is 3 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2518   +/-   ##
=======================================
  Coverage   99.98%   99.98%           
=======================================
  Files         193      193           
  Lines       16942    16944    +2     
=======================================
+ Hits        16940    16942    +2     
  Misses          2        2           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[facebook-github-bot]

This pull request has been merged in b58852e.