refactor: build InfiniLM from the modern Infini stack

[voltjia]

Summary

• Build the NVIDIA InfiniLM image from the caller checkout and the InfiniRT, InfiniOps, and InfiniCCL revisions pinned by InfiniCore.
• Fetch Core and InfiniLM third-party sources at exact revisions with shallow, retryable operations and remove the legacy Core build path.
• Pass the caller project root to xmake setup, add focused workflow and Dockerfile contracts, and preserve the shared stack environment in interactive shells.

Motivation

The previous image assembled the inference stack from duplicated legacy sources and left interactive shells pointed at obsolete install and CUTLASS paths. The shared CI image now follows the same component ownership and installed-prefix contract as the production build.

Validation

• Passed 10/10 tests.test_infinilm_ci contract tests.
• Passed Ruff check, Ruff format check, and git diff --check.
• Built images/nvidia/Dockerfile.deploy from clean GitHub checkouts with Core 5b2fe0a4, InfiniLM f9a2717, and this branch at 0e070166.
• Verified the resulting CUDA 13.1 image on NVIDIA A100 with cross-device copy, InfiniRT graph capture/replay, Qwen3-0.6B TP1, and Qwen3-0.6B TP2.
• Verified an interactive shell keeps INFINI_ROOT=/opt/infini-stack/prefix, imports the installed packages, and does not set the removed CUTLASS environment.

Scope

This PR targets the infiniCore_ci integration branch. Validation for non-NVIDIA images is intentionally deferred to the later cross-platform phase.