drop support for maximum and minimum

[vtavana]

drop support for maximum and minimum for both float and complex dtypes since they do not bring performance improvement compared to stock NumPy implementation.

timing on: Intel(R) Xeon(R) Platinum 8480+ OpenCL 3.0 (Build 0) [2025.20.7.0.08_160000.prerelease]

on the main branch:
Function maximum:

Function: maximum | Dtype: float64
      Size |   NumPy (ms) |  MKL_umath (ms)
---------------------------------------------
     10**4 |        0.003 |           0.039
     10**5 |        0.053 |           0.514
     10**6 |        0.730 |           5.165
     10**7 |       22.584 |          62.269
     10**8 |      279.800 |         647.993

Function: maximum | Dtype: complex128
      Size |   NumPy (ms) |  MKL_umath (ms)
---------------------------------------------
     10**4 |        0.041 |           0.087
     10**5 |        0.522 |           0.877
     10**6 |        4.829 |           8.758
     10**7 |       74.019 |         113.827
     10**8 |      751.199 |        1150.744

Function minimum:

Function: minimum | Dtype: float64
      Size |   NumPy (ms) |  MKL_umath (ms)
---------------------------------------------
     10**4 |        0.003 |           0.038
     10**5 |        0.051 |           0.515
     10**6 |        0.929 |           5.153
     10**7 |       22.894 |          62.631
     10**8 |      278.826 |         647.707
     
Function: minimum | Dtype: complex128
      Size |   NumPy (ms) |  MKL_umath (ms)
---------------------------------------------
     10**4 |        0.042 |           0.083
     10**5 |        0.489 |           0.852
     10**6 |        4.854 |           8.416
     10**7 |       73.240 |         108.444
     10**8 |      764.019 |        1125.169 

[ndgrigorian]

considering the significant slowdown this LGTM

I wonder if this performance issue should be passed to MKL team

[vtavana]

considering the significant slowdown this LGTM

I wonder if this performance issue should be passed to MKL team

MKL does not have any equivalent function for maximum and minimum and mkl_umath implementation was based on primitive/built-in operators. This implementation is similar to what was used in numpy-1.22. However, stock numpy-1.23, the implementation was updated which is faster and so current numpy is faster than mkl_umath.

MKL only has equivalent function for float loops of fmax and fmin which are already been used in mkl_umath.
And I am going to drop support for complex loop of fmax and fmin as well since there is no performance improvement there.