Add custom COCO dataset loader function

[lumurillo]

Use PyTorch to load custom COCO compatible datasets into Armory.

Example of scenario configuration JSON:

diff --git a/scenario_configs/eval7/carla_overhead_object_detection/carla_obj_det_adversarialpatch_undefended.json b/scenario_configs/eval7/carla_overhead_object_detection/carla_obj_det_adversarialpatch_undefended.json
index cb914461..4f271427 100644
--- a/scenario_configs/eval7/carla_overhead_object_detection/carla_obj_det_adversarialpatch_undefended.json
+++ b/scenario_configs/eval7/carla_overhead_object_detection/carla_obj_det_adversarialpatch_undefended.json
@@ -6,7 +6,7 @@
         "kwargs": {
             "batch_size": 1,
             "learning_rate": 0.05,
-            "max_iter": 500,
+            "max_iter": 10,
             "optimizer": "Adam",
             "targeted": false,
             "verbose": true
@@ -17,11 +17,10 @@
     },
     "dataset": {
         "batch_size": 1,
-        "eval_split": "dev",
-        "framework": "numpy",
-        "modality": "rgb",
-        "module": "armory.data.adversarial_datasets",
-        "name": "carla_over_obj_det_dev"
+        "dataset_dir": "/home/luismr/carla-datagen-toolkit/data/dev_eval7/rgb",
+        "ann_file": "/home/luismr/carla-datagen-toolkit/data/dev_eval7/kwcoco_annotations.json",
+        "module": "armory.data.datasets",
+        "name": "custom_coco_dataset"
     },
     "defense": null,
     "metric": {
@@ -53,7 +52,8 @@
     "scenario": {
         "kwargs": {},
         "module": "armory.scenarios.carla_object_detection",
-        "name": "CarlaObjectDetectionTask"
+        "name": "CarlaObjectDetectionTask",
+        "export_batches": true
     },
     "sysconfig": {
         "docker_image": "twosixarmory/armory",

Command used to test this feature:

CUDA_VISIBLE_DEVICES=1 armory run --no-docker --skip-attack --use-gpu --num-eval-batches 1 scenario_configs/eval7/carla_overhead_object_detection/carla_obj_det_adversarialpatch_undefended.json