Make Inputs handle non-aligned and non-padded buffer

[charles-paperman]

For some buffer, it is impossible to padd and to aligned without performing a full copy.
The inputs layer could:

• Detect that the start address is not aligned, build a custom first block to overpass the alignement padding artificially the begining with null byte
• Detect that the end address is not aligned, build a custom last block in the same spirit.

Additional context
This is useful when you don't have any control on the origin of the data: python or when your data is a substring of something else and you want to avoid copying.

[github-actions]

Tagging @V0ldek for notifications

[V0ldek]

This is doable, but overhead has to be measured.

We already have proof-of-concept code for handling the last block separately when it comes to padding. We could split off the first block and treat it separately as well. The question is how big of a performance impact this would have on small inputs, so we need some small benches to test this.

[V0ldek]

Released in v0.8.4