docker: install transformer_engine_torch runtime deps

[yueming-yuan]

Problem

#1796 switched TransformerEngine to the rolling wheels, installed with
--no-deps. Pinning the three TE dists to exactly those wheels is the point of
that flag, but it also drops transformer_engine_torch's own runtime
requirements:

torch>=2.1, einops, onnxscript, onnx, packaging, pydantic,
nvdlfw-inspect, transformer_engine_cu13==2.17.0

transformer_engine.pytorch imports onnxscript unconditionally on its core
import path:

transformer_engine/pytorch/__init__.py:21   -> module/__init__.py
  module/layernorm_linear.py:21 -> module/base.py:23 -> module/_common.py:15
    export.py:61 -> onnx_extensions.py:26 -> import onnxscript

so every image built since #1796 dies as soon as anything touches TE — which is
every GPU test:

File ".../transformer_engine/pytorch/onnx_extensions.py", line 26, in <module>
    import onnxscript
ModuleNotFoundError: No module named 'onnxscript'

Seen on #1795's stage-c-*-gpu-h200 jobs (run 30190570673), where it takes out
the run before any test body executes. It is not specific to that PR — it
reproduces on any image built from current main.

Why the build stayed green

verify_transformer_engine.py compared versions and metadata and then called

importlib.util.find_spec("transformer_engine")

find_spec resolves the package directory without importing it, so it succeeds
against an installation that cannot actually be imported. The verifier passed,
the image was pushed, and the failure surfaced hours later in GPU tests.

Change

• Install the non-TE runtime deps after the pinned --no-deps triplet.
• Have the verifier assert that every non-TE requirement of
  transformer_engine_torch resolves, so a future --no-deps gap fails at the
  layer that introduced it rather than shipping a green image.

Deliberately not dropping --no-deps: letting pip resolve the triplet is what
that flag exists to prevent.

Verification

Ran the new verifier inside a miles image that has the deps, then hid
onnxscript's dist-info to reproduce exactly the state #1796 produces:

--- deps present (expect PASS) ---
PASS
--- simulate the #1796 state: hide onnxscript ---
AssertionError: transformer_engine_torch runtime deps not installed: ['onnxscript']
--- restored ---
PASS again

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[guapisolo]

Is it possible to add this file to CI?

[guapisolo]

LGTM. wait for CI green and e2e pass then we can merge.

[Zhichenzzz]

Looks great! should we merge it now?