fallback to scipy when linalg.matrix_power isn't available

[timmysilv]

Context:
Tensorflow does not have a matrix_power function. For non-int powers, we always fall back to scipy, and we noticed that this works with Tensorflow as well! We can use this...

Description of the Change:

• Whenever trying to raise TF matrices to an integer matrix-power, use scipy.linalg.fractional_matrix_power. Until today, we just raised an error. Pow is not trainable regardless, so there's no harm in adding this.
• This also indirectly allowed for raising any QubitUnitary to a non-int power (eg: qml.QubitUnitary(np.eye(2), wires=[0]).pow(3.5))

Benefits:
Less errors! more features!

Possible Drawbacks:

• Still not differentiable
• QubitUnitary.pow looks more complex to support more stuff

Related GitHub Issues:
Fixes #4825 [sc-49952]

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (456d98e) 99.66% compared to head (47c33d6) 99.65%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #4827      +/-   ##
==========================================
- Coverage   99.66%   99.65%   -0.01%     
==========================================
  Files         382      382              
  Lines       34468    34213     -255     
==========================================
- Hits        34351    34095     -256     
- Misses        117      118       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[lillian542]

🚀