Improve newer-GPU compatibility and optional kernel fallbacks

[Rajesh-Nimmalapudi]

Summary

This PR improves IntelliFold compatibility on newer NVIDIA GPUs while keeping
the default inference path unchanged for existing users.

The changes in this PR:

• relax the hard torch==2.6.0 pin so users can install a PyTorch build that
  supports newer GPUs
• move deepspeed from a mandatory dependency to an optional extra
• keep baseline inference available when DeepSpeed is installed but cannot be
  imported
• fall back to the standard attention path when DS4Sci is explicitly requested
  but unavailable
• remove hardcoded older CUDA arch flags from the fast_layernorm extension
  build
• document a clearer installation path for newer GPU architectures and clarify
  that DS4Sci and fast_layernorm are optional kernel paths

Motivation

The current packaging and kernel-build defaults make installation on newer
NVIDIA GPUs harder than necessary:

• strict torch pinning prevents users from selecting a newer torch/CUDA build
  for newer devices
• deepspeed is treated as mandatory even though baseline inference does not
  require it
• fast_layernorm hardcodes an older CUDA architecture set

This PR keeps the baseline path conservative while removing those avoidable
compatibility blockers.

What changed

Packaging

• setup.py
  • changed torch==2.6.0 to torch>=2.6.0
  • moved deepspeed into extras_require["deepspeed"]
• environment.yaml
  • changed torch==2.6.0 to torch>=2.6.0
  • removed mandatory deepspeed

Runtime behavior

• intellifold/openfold/model/primitives.py
  • treat DeepSpeed as optional at import time
• intellifold/openfold/utils/kernel/deepspeed_compat.py
  • centralize DS4Sci availability checks
• intellifold/openfold/inference_config.py
• intellifold/openfold/v2_inference_config.py
• intellifold/openfold/v2_flash_inference_config.py
  • disable DS4Sci cleanly when it is requested but unavailable
  • keep the standard path active instead of carrying a broken DS4Sci request
    into model execution

Kernel build behavior

• intellifold/openfold/utils/layer_norm/torch_ext_compile.py
  • removed hardcoded TORCH_CUDA_ARCH_LIST="7.0;8.0"
  • removed the fixed sm_70/sm_80/sm_86/sm_90 list
  • respect TORCH_CUDA_ARCH_LIST when it is provided
  • otherwise defer to PyTorch's default extension-build behavior

Documentation

• added a short newer-GPU installation note to README.md
• added an explicit newer-GPU install path to docs/installation.md
• clarified kernel build expectations in docs/kernels.md

Validation

I validated the branch locally with:

• baseline runtime on a newer NVIDIA GPU using a newer torch/CUDA stack
• fast_layernorm JIT build and runtime on the same GPU
• DS4Sci-request fallback behavior from a clean installed package
• package build and install from this branch into a clean target directory

Scope

This PR does not make DS4Sci the default path. DS4Sci remains optional, and
this change is focused on compatibility and safer optional-kernel behavior.