Add general lazy evaluation for all blosc2 functions

[lshaw8317]

Although there was no double evaluation in old versions of blosc2 for reductions as there was an is_inside_new_expr() stack check which aborted the evaluation and returned blosc2.zeros during the eval call, we have added a shape parser to avoid having to arrive at reduce_slices for reductions. In addition, this allows us to handle other non-numexpr functions without adding the same is_inside_newexpr() call.
This PR is thus useful since 1) it adds the possibility to compute with arbitrary blosc2 functions (such as matmul or even logaddexp) in a lazy imperative way via the changes to ne_evaluate and 2) adds a reasonably extensive shape parser function, which is a bit more explicit than the returning blosc2.zeros hack.

For the following benchmark:

import time
import blosc2

# --- Experiment Setup ---
N = 100000
X = blosc2.ones((N,N))
t0 = time.time()
blosc2.evaluate('sum(X)')
dt = time.time()
print(f"Time for blosc2.evaluate('sum(X)'):{round(dt, 3)} s")

one can confirm that the new method also evaluates the reduction only once on the full operands, and takes the same time as in blosc2<=3.9.1.

[FrancescAlted]

I only have a couple of comments, but mainly cosmetic.

Also, make sure that the new way to compute, e.g. reductions, does perform similarly to the previous one.