fix(train): write a sourceable sm_training.env for quotes and non-identifier names

[AkshayShah03]

Issue #, if available: N/A

Description of changes:

The training container's script runs set -e with an ERR trap and sources /opt/ml/input/sm_training.env. set_env writes every variable as export NAME='value', which breaks in two cases:

• A single quote in any value breaks the file's quoting. A hyperparameter like it's also ends up in SM_HPS and SM_TRAINING_ENV:

  sm_training.env: line 26: unexpected EOF while looking for matching `''
  

• A channel name or hyperparameter key containing - or ., both allowed by the API (e.g. channel train-data, key model.lr), produces an invalid identifier:

  export: `SM_CHANNEL_TRAIN-DATA=/opt/ml/input/data/train-data': not a valid identifier
  

Either one kills the job before the entry script runs.

This change maps non-identifier characters in names to _; SM_HP_* already did this for -. It also escapes single quotes in values with the standard '"'"' sequence. The output for ordinary values is unchanged; the existing byte-for-byte test_set_env still passes. The fix is applied to both the sagemaker-train and the sagemaker-core copies of environment.py.

Testing:

• A new test in each package writes the file with a train-data channel, a model.lr key and an it's a test value, then sources it with bash -c 'set -e; trap ... ERR; source ...' and checks the values. It fails before this change.
• container_drivers tests: 33 passed (train) and 79 passed (core).

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.