Clean up dockerfile and use devel base image

[grzegorz-roboflow]

Description

Update based on progress made by @PawelPeczek-Roboflow

• aligned image with cuda 11.8 - instead of previously attempted 11.7 - since we explicitly require torch cu118 - and got rid of all attempts to install / uninstall / re-install cudnn which turned out not to be a problem after switch
• removed requirements/requirements.pali.local.txt and put all requirements into requirements/requirements.pali.server.txt, as this model only runs at server
• made sure that requirements are compliant with requirements from other models - including SAM, Grounding Dino and CogVLM - for SAM we theoretically were allowing for old torch 1.0, and now it is no longer possible, but this is fair trade since torch 2.0 is now standard
• below you can see diff of packages from build v0.13.0 and current:

  [('torchvision', '0.15.2', '0.15.2+cu118'),
  ('filelock', '3.13.1', '3.15.4'),
  ('openai', '1.35.5', '1.35.12'),
  ('botocore', '1.34.133', '1.34.142'),
  ('rapidfuzz', '3.9.3', '3.9.4'),
  ('time-machine', '2.14.1', '2.14.2'),
  ('tifffile', '2024.6.18', '2024.7.2'),
  ('jsonschema', '4.22.0', '4.23.0'),
  ('pydantic_core', '2.18.4', '2.20.1'),
  ('orjson', '3.10.5', '3.10.6'),
  ('sympy', '1.12', '1.13.0'),
  ('awscli', '1.33.15', '1.33.24'),
  ('pydantic', '2.7.4', '2.8.2'),
  ('torch', '2.0.1', '2.0.1+cu118'),
  ('ultralytics', '8.2.45', '8.2.52'),
  ('fonttools', '4.53.0', '4.53.1'),
  ('cmake', '3.29.6', '3.30.0'),
  ('google-api-python-client', '2.134.0', '2.137.0'),
  ('readme_renderer', '43.0', '44.0'),
  ('fsspec', '2024.2.0', '2024.6.1'),
  ('ipykernel', '6.29.4', '6.29.5'),
  ('mplcursors', '0.5.3', '0.5.2'),
  ('Jinja2', '3.1.3', '3.1.4'),
  ('pillow', '10.3.0', '10.4.0'),
  ('ipython', '8.25.0', '8.26.0'),
  ('google-auth', '2.30.0', '2.32.0'),
  ('nh3', '0.2.17', '0.2.18'),
  ('certifi', '2024.6.2', '2024.7.4'),
  ('rasterio', '1.3.0', '1.3.10'),
  ('matplotlib', '3.9.0', '3.9.1'),
  ('rpds-py', '0.18.1', '0.19.0')]

• changes ^ are minor or not relevant
• Turned out that transformers>=4.42.0 breaks CogVLM!
• flash_attn was needed to be in separate requirements file as it requires other dependencies to be already installed, rather than installed alongside with this package in single pip install
• Final image has 19.4GB which is over +10GB from last build
  • Tried lean build from runtime image - but then for flash_attn we need nvcc compiler which comes along with cuda toolkit - once installed - dependency expolded to size near to 30GB - bigger than starting from devel
  • also attempted to do multi-stage build properly (see commit) - basically using devel image as starting point, then switching to runtime - but I could only save 5GB from 19.4GB and despite things building - it was complaining on tensorrt so library - which was not present in devel when recursively searched for - so must have been built during flash_attn build probably - not sure where it sits and could not copy it and place properly into target image
• added CI action to verify install of all extras for inference and inference-gpu on each commit to PR branch to main
• added integration tests of GPU server for all LMMs
• fixed SAM integration tests
• made integration tests running on each merge to main
• added test CI action to verify behaviour of conditioning in release action!
• @probicheaux I must have changed florence implementation as it was broken by attention masks being passed to generate() - not sure if thats due to specific transformers version, but the code is compliant in behaviour with this: https://huggingface.co/microsoft/Florence-2-large
• I removed Dockerflie paligemma as it is no longer needed - base docker GPU build has it installed

This is not done, but @probicheaux please update docs for using paligema and florence - when we have release, rel notes would have relevant info

List any dependencies that are required for this change.

Type of change

Please delete options that are not relevant.

• [x] Bug fix (non-breaking change which fixes an issue)
• [x] New feature (non-breaking change which adds functionality)
• [ ] This change requires a documentation update

How has this change been tested, please provide a testcase or example of how you tested the change?

• extened tests
• GPU regression tests work: https://github.com/roboflow/inference/actions/runs/9874436434
• Jetson 5 local build works

Any specific deployment considerations

For example, documentation changes, usability, usage/costs, secrets, etc.

Docs

• [ ] Docs updated? What were the changes:

[probicheaux]

Ok, this is finally consistent

@PawelPeczek-Roboflow , sorry I didn't get to integration tests in time, and I'm out for a week starting tomorrow @hansent , this is looking good to go locally

[probicheaux]

We can probably remove the paligemma Dockerfile too