Good to hear that, can you provide a pull request and corresponding benchmark test case?

Are you saying that I should submit a PR with benchmark code or also with the code I suggested above? If the latter, there are probably hundreds of ways to implement it and the code above is just the first thing that came to my mind; so why use exactly that code?

As an implementation idea to improve performance, I prototyped a "block-wise" trim function to be used before trim_zeros

How about converting the passed object into a boolean array and then use np.argmax() to find the first/last non-zero element?
With your previously defined example array I'm seeing an increase in execution speed of ~2 orders of magnitude (398 µs versus 37 ms).

import numpy as np

def trim_zeros(filt, trim='fb'):
    a = np.asanyarray(filt, dtype=bool)
    if a.ndim != 1:
        raise ValueError('trim_zeros requires an array of exactly one dimension')

    trim_upper = trim.upper()
    len_a = len(a)
    i = j = None
    
    if 'F' in trim_upper:
        i = a.argmax()
        if not a[i]:  # i.e. all elements of `filt` evaluate to `False`
            return filt[len_a:]

    if 'B' in trim_upper:
        j = len_a - a[::-1].argmax()
        if not j:  # i.e. all elements of `filt` evaluate to `False`
            return filt[len_a:]

    return filt[i:j]

Does that code work without the if not ...s?

Does that code work without the if not ...s?

Without the if not ... it will fail if the input array consists entirely of zeros,
in which case argmax() will always return 0 and thus the filt[i:j] == filt.

>>> import numpy as np

>>> a = np.zeros(10, dtype=bool)
>>> i = a.argmax()
>>> j = len(a) - a[::-1].argmax()

>>> print(i, j)
0 10

>>> print(np.all(a == a[i:j]))  # Uhoh, `a` is not being trimmed
True

Another option is to check with np.any() right at the beginning, though this appears to be a bit slower.

in which case argmax() will always return 0

Ah, I thought it might return len(a) - 1

Shall I create a pull request with the implementation as proposed above?

I've just created a pull request for the issue at #16911.