The noise floor is the random variation added by the sensor and especially its readout electronics. In practice, black-level subtraction removes a fixed offset, not random noise.

Think of black level as a baseline added so the signal stays above zero. Subtracting that baseline later only shifts the whole signal downward; it does not remove the fluctuations riding on top of it. If the baseline is near the noise floor, values below it are effectively lost, but the noise itself still exists.

Why not subtract it in analog before the ADC? Because the analog circuitry that performs that subtraction is also noisy. Amplifiers, op-amps, and related circuits introduce their own noise, and higher gain makes that noise more visible. A DC-cancellation or black-level removal stage can reduce offset, but high-frequency noise from the sensor/readout chain still passes through, and the subtraction stage adds some noise of its own.

So black-level subtraction is useful for correcting offset, but it cannot completely eliminate the noise floor, because the floor is dominated by random electronic noise, not just a fixed black value.