Pretrained weights without a ".weights" suffix are silently ignored

[pete-machine]

To report a bug, please provide the following:

• the output of the darknet version command

Darknet v2.0-208-ga7901e02-dirty
CUDA runtime version 12010 (v12.1), driver version 12020 (v12.2)
cuDNN version 12020 (v8.9.5), use of half-size floats is ENABLED
=> 0: Tesla T4 [#7.5], 14.6 GiB
OpenCV v4.2.0

• the exact command you ran
  Something like this:

./darknet detector train data/obj.data yolo-obj.cfg yolov4.conv.137

• the operating system you are using
  Ubuntu 22.04

I found that pre-trained weights without the ".weights" suffix are silently ignored in this fork of darknet.
This is a critical issue for users transitioning from AlexeyAB, where pretrained-models or custom model weights
are listed without a .weights suffix (AlexeyAB documentation)... (Instead model files uses the final layer as a suffix e.g. 137 for yolov4.conv.137 )

Users may incorrectly assume that pre-trained weights are loaded with HankAI and they will either 1) achieve worse performance without ever noticing or 2) wonder why the model converges a lot slower than AlexeyAB - at least I did.

Showing hankai and Alexey trainings with and without pretrained weights

The line for mapping command line arguments to weight filename is:

# From line 146 in darknet_cfg_and_state.cpp
if (extension == ".weights"	and weights_filename.empty())	{ weights_filename	= path; }

A fix could be to throw an error if all arguments are not parsed correctly or throw an error message if a user tries to load a model with a integer-suffix.