Remove unused resize_mode in Torch path

The resize mode is no longer used by the test.lua Torch wrapper.
Besides this field isn't part of the common dataset interface.
Using it is causing problems during multiple image inference in Torch when using generic datasets.
This should have been changed in NVIDIA#723

digits/model/tasks/torch_train.py

@@ -807,7 +807,6 @@ def infer_many_images(self, images, snapshot_epoch=None, gpu=None):
                     '--testMany=yes',
                     '--allPredictions=yes',   #all predictions are grabbed and formatted as required by DIGITS
                     '--image=%s' % str(temp_imglist_path),
-                    '--resizeMode=%s' % str(self.dataset.resize_mode),   # Here, we are using original images, so they will be resized in Torch code. This logic needs to be changed to eliminate the rework of resizing. Need to find a way to send python images array to Lua script efficiently
                     '--network=%s' % self.model_file.split(".")[0],
                     '--networkDirectory=%s' % self.job_dir,
                     '--snapshot=%s' % file_to_load,

tools/torch/test.lua

@@ -22,7 +22,6 @@ require 'Optimizer'
 --print 'processing options'
 
 opt = lapp[[
--m,--resizeMode (default squash) Resize mode (squash/crop/fill/half_crop) for the input test image, if it's dimensions differs from those of Train DB images.
 -t,--threads (default 8) number of threads
 -p,--type (default cuda) float or cuda
 -d,--devid (default 1) device ID (if using CUDA)