.npy files are very useful for exchanging numeric arrays between Python and C.

The main motivating cases for including this in emlearn are:

• Checking/testing of the on-device pipeline and models, using a validation dataset (exported from Python on host)
• Passing data when validating converted models on the host. Will eventually replace eml_test_read_csv used in emlearn.common.CompiledClassifier
• Support feature extraction to enable C feature extraction to be used transparently from Python (future)

This implementation aims to support constrained microcontrollers. This means:

• Support streaming processing of data in chunks.
• No dynamic allocation.
• Small in terms of code size
• Work with any file/stream I/O abstraction

Conceptually it is similar to https://github.com/jonnor/micropython-npyfile (for MicroPython)