This commit improves the performance slightly of the
OneQubitEulerDecomposer by avoiding the use of append(), which does extra
validation of the arguments when constructing the output circuits.
Instead the _append() method is used which does not have the validation
overhead, however this is not necessary since we know how the circuit is
constructed and the arguments when adding the gate are being set
correctly (there is only 1 qubit anyway).

The np.isclose() function quickly becomes a bottlneck in the
OneQubitEulerDecomposer class as the majority of the operations are
comparing angles to decide when to insert gates. However, we can avoid
this overhead by switching from numpy's isclose function which is
optimzed for arrays not floats [1][2] to just using the standard
library's math module's isclose() function. This commit makes this
change as well as using intermediate variables to avoid duplicate
calculations to speed up the decomposer.

According to the np.mod() docs [1] the np.mod() is equivalent to the
stdlib '%' operator in python however it is designed to work with numpy
array's instead of single values. This adds about an order of magnitude
overhead to the mod operations (on the order of 1us vs 100ns) which can
add up as _mod2pi is called multiple times. However python's mod
operator is not ideal for working with floats [2] and when used produces
different results than expected. To avoid this but still improve
performance this commit switches to use the stdlib math module's
fmod() function, which produces the expected result and is only
marginaly slower than the '%' operator, which will still avoid the
overhead of the numpy function.

[1] https://numpy.org/doc/stable/reference/generated/numpy.mod.html
[2] https://docs.python.org/3/library/math.html#math.fmod