fix: add dedicated glaucoma detection endpoint with separate model

[Siddh2024]

Fixes #56

Description

Added a fully separate glaucoma detection pathway to prevent the glaucoma endpoint from loading cataract-trained or brain-tumor model weights. The glaucoma model is loaded from its own dedicated file path (glaucoma_model.h5 by default, configurable via GLAUCOMA_MODEL_PATH env var) and never shares weights with any other condition model.

Changes

• New endpoint: POST /api/predict/glaucoma - accepts fundus images, runs glaucoma-specific preprocessing (CLAHE-enhanced LAB), and returns glaucoma/normal scores
• New model loading: glaucoma_model global loaded independently of the brain tumor ensemble; endpoint returns 503 if model file is absent
• New preprocessing: preprocess_fundus_image() with CLAHE and fundus-appropriate normalization
• Config: GLAUCOMA_MODEL_PATH env var (defaults to glaucoma_model.h5), GLAUCOMA_CONFIDENCE_THRESHOLD config
• Documentation: Updated .env.example with glaucoma model path docs
• Stats endpoint: Reports glaucoma model availability
• Startup logs: Shows glaucoma model status on boot

Why this fixes #56

The original issue stated the glaucoma endpoint loaded catract_model.pth - the same weights used for cataract predictions. This PR ensures the glaucoma endpoint uses a dedicated model loaded from its own file path, making predictions meaningful and condition-specific.

[Brijeshthummar02]

@Siddh2024 go through the weights available in discussion page and show me logs after tests of this addition of glaucoma model on scans.

[Brijeshthummar02]

it's confusing to me why we adding it fully separate model even we can add this into existing model in combination of resnet-50. what's POV behind separate work?