ENH: Deprecate ``np.squeeze()`` with unspecified ``axis``

Proposed new feature or change:

Deprecate calls unspecified axis in calls to np.squeeze(), and/or issue a single runtime warning.

Rationale

squeeze() without an axis argument is the cause of many subtle bugs. Example:

x = images[:batch_size]

B, C, H, W = x.shape

# Incorrect attempt to remove "grayscale" channel dimension.
gray = x.squeeze()

This causes hard-to-debug issues if:

• B == 1 (batch size is 1)
• H == 1 (height is 1)
• W == 1 (width is 1)

The above should have been written:

gray = x.squeeze(1)

Examples in the wild

Many users of this library are not "experts" at code, and frequently produce and publish programs containing this bug.

Roughly 486000 bugs here:
https://github.com/search?q=language%3APython+squeeze%28%29&type=code
I challenge the reader to find an example which is not a (potential) bug!

YOLOX (10k stars):

        conf_mask = (image_pred[:, 4] * class_conf.squeeze() >= conf_thre).squeeze()

The correct code would have been:

        conf_mask = image_pred[:, 4] * class_conf.squeeze(1) >= conf_thre

The above bug actually inspired the creation of this ENH proposal.

Start at the np.squeeze() entry point and locate its implementation and related tests; no specific files or tests are named in the issue. Decide whether unspecified axis should be deprecated or produce a single runtime warning, then add coverage showing the intended behavior for the documented shape cases.