Simplify tensor operations and interpolation

[mspronesti]

Hi, this PR aims at simplifying some of the operations done in the setup of some of the schedulers I implemented in the past weeks, relying more on the tch crate and reducing the number of casts to vectors.

Specifically:

• simplify tensor interpolation: previously, since I didn't know tch supported a pytorch-equivalent of ge, called ge_tensor - ge in tch only accepts a scalar - I used a vector of tensors whose elements were eventually stacked. Now, I only work with tensors, thus porting exactly the python version.
• reduce number of casts from Tensors to Vecs, mainly relying more on cat and linspace instead of working with ranges and Vec::push
• stylistic change: wherever possible, I replaced slice with .i(<range>) to improve readability

[LaurentMazare]

Looks nicer indeed, thanks!