Improve the performance of removing variables from expressions

[arcondello]

This provides a pretty significant performance improvement when fixing variables in individual constraints.

import time

import dimod

num_variables = 10_000

cqm = dimod.ConstrainedQuadraticModel()
cqm.add_variables("BINARY", num_variables)
c0 = cqm.add_constraint(((v, 1) for v in range(num_variables)), '==', 1)
expr = cqm.constraints[c0].lhs

t = time.perf_counter()
for v in range(num_variables):
    expr.fix_variable(v, 0)
print(time.perf_counter() - t)

Gives a time of ~0.15s vs 1.37s on 0.12.4 on my system.

The majority of the improvement comes from optimizing remove_variable(). In fact, the C++ fix_variable() is not even called from the Python code. I did test adding a dedicated Cython call rather than using the view, but that doesn't change the performance very much at all since the vast majority of the time is spent on rebuilding the indices.

Additional Context
This is a first step to addressing #1317. But it doesn't actually touch the relevant code (yet).

[randomir]

LGTM!