[trainer/deepspeed] load_best_model (reimplement re-init) (#17151)

* [trainer/deepspeed] load_best_model

* to sync with DS PR #1947

* simplify

* rework load_best_model test

* cleanup

* bump deepspeed>=0.6.5

Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>

setup.py

@@ -27,7 +27,7 @@ To create the package for pypi.
 
 3. Unpin specific versions from setup.py that use a git install.
 
-4. Checkout the release branch (v<RELEASE>-release, for example v4.19-release), and commit these changes with the 
+4. Checkout the release branch (v<RELEASE>-release, for example v4.19-release), and commit these changes with the
    message: "Release: <VERSION>" and push.
 
 5. Wait for the tests on main to be completed and be green (otherwise revert and fix bugs)
@@ -103,7 +103,7 @@ _deps = [
     "cookiecutter==1.7.3",
     "dataclasses",
     "datasets",
-    "deepspeed>=0.6.4",
+    "deepspeed>=0.6.5",
     "dill<0.3.5",
     "fairscale>0.3",
     "faiss-cpu",