Fix best_dev_checkpoint path

[wasertech]

Produced checkpoints are using absolute path where they should really be using relative ones.

Checkpoints FR STT 0.9

# ./best_dev_checkpoint
model_checkpoint_path: "/mnt/checkpoints/best_dev-221133"
all_model_checkpoint_paths: "/mnt/checkpoints/best_dev-221133"

Here is English checkpoint for reference.

# ./best_dev_checkpoint
model_checkpoint_path: "best_dev-3663881"
all_model_checkpoint_paths: "best_dev-3663881"

Fix should probably go into package.sh I think.

Related : coqui-ai/STT#2338
^ STT's lm_optimizer throws a NotFoundError from TF, since the path points to nowhere in transfer-learning configuration, as checkpoint path is /transfer-checkpoint/ instead of /mnt/checkpoints/.

commonvoice-fr/DeepSpeech/CONTRIBUTING.md

Lines 76 to 78 in 5699e59

	To use an existing checkpoint, just ensure the `docker run` includes a mount such as:
	`type=bind,src=PATH/TO/CHECKPOINTS,dst=/transfer-checkpoint`. Upon running, data
	will be copied from that place.

commonvoice-fr/DeepSpeech/train.sh

Lines 14 to 17 in 5699e59

	if [ -f "/transfer-checkpoint/checkpoint" -a ! -f "/mnt/models/output_graph.pb" ]; then
	echo "Using checkpoint from ${TRANSFER_CHECKPOINT}"
	cp -a /transfer-checkpoint/* /mnt/checkpoints/
	fi;

That was for DS but with STT, we have access to --load_checkpoint_dir and --save_checkpoint_dir flags instead.

	if [ -f "/transfer-checkpoint/checkpoint" -a ! -f "/mnt/models/output_graph.tflite" ]; then
		echo "Using checkpoint from ${TRANSFER_CHECKPOINT}"
		# use --load_checkpoint_dir for transfer learning
		LOAD_CHECKPOINT_FROM="--load_checkpoint_dir /transfer-checkpoint --save_checkpoint_dir /mnt/checkpoints"
	else
		LOAD_CHECKPOINT_FROM="--checkpoint_dir /mnt/checkpoints/"
	fi;

Example from my branch stt140-cv9