[v2 BUG]: Non integer values for 'batch-size' in hpopt output

[akshatzalte]

Performing hyper-parameter optimization on 'all' hyper-parameters gave a non-integer value for the batch size. Due to some bug, the search space is continuous rather than discrete for batch size.

Here is the input hpopt.sh file:
`#!/bin/bash -l

conda activate chemprop

results_dir="."
data_path="/home/akshatz/bond_order_free/data_2_kcal_def/dataset/run1/data_run_1.csv"
split_path="/home/akshatz/bond_order_free/data_2_kcal_def/dataset/run1/splits.json"

chemprop hpopt
-t regression
--data-path $data_path
--splits-file $split_path
--hyperopt-n-initial-points 25
--raytune-num-samples 50
--epochs 100
--raytune-grace-period 100
--hyperopt-random-state-seed 12
--aggregation sum
--search-parameter-keywords all
--num-workers 20
--hpopt-save-dir $results_dir
--smiles-columns SMILES
--target-columns H298_kcal
--add-h
--keep-h
--raytune-use-gpu`

Here are the contents of the best_parameters.json file generated:
{
"train_loop_config": {
"activation": "TANH",
"final_lr_ratio": 0.009012526078574787,
"message_hidden_dim": 2200.0,
"max_lr": 8.85525109510601e-06,
"batch_size": 21.763120334996223,
"aggregation_norm": 23.0,
"depth": 3.0,
"ffn_num_layers": 2.0,
"dropout": 0.0,
"ffn_hidden_dim": 1100.0,
"init_lr_ratio": 0.00032278162414634416,
"aggregation": "mean",
"warmup_epochs": null
}
}