Mobile responsive fixes, filter schema row & fix broken A2A link

[Chesars]

Summary

A pass over the mobile experience plus two correctness fixes for upstream data the page consumes.

Bugs fixed

• sample_spec was rendering as a real model. The upstream model_prices_and_context_window.json ships a sample_spec pseudo-entry where each field's value is a description string for that field (e.g. mode: \"one of: chat, embedding, ...\", max_input_tokens: \"max input tokens, if the provider specifies it...\", litellm_provider: \"one of https://docs.litellm.ai/docs/providers\"). It was being treated as a real row, breaking the layout: the mode-badge ballooned to fit the schema string, the Context column literally printed \"max input tokens, if the provider...\", and the row pushed the table wider than the viewport so the Model column header appeared to overlap the data. Filter it out at load time.
• A2A provider doc link 404'd. Upstream JSON has https://docs.litellm.ai/docs/providers/a2a (404) but the actual page lives at https://docs.litellm.ai/docs/a2a. Added a small per-provider URL override map in Providers.svelte.

Mobile responsive pass

• Removed position: sticky from <thead> — its top: 63px was causing rows to scroll behind the column headers, which read as overlap (especially after expanding a sample-spec panel on a phone).
• Detail panel reflow: section headings (Pricing, Model Info, Features) sized down a tier, pricing card labels/values reduced for hierarchy, info-rows now stack label-over-value so long values like 1,000,000 tokens can't collide with their label.
• Code snippets (Python SDK / AI Gateway proxy): pre-wrap + word-break: break-word so long model names like bedrock/.../nova-canvas-v1:0 wrap instead of forcing a visually huge horizontally-scrolling block. Tablet breakpoint added for an intermediate size between desktop and phone.
• Locked model-row height (52 px desktop / 48 px mobile) so Bedrock and the first entries don't render visibly taller than the rest.

Trust-logo hover unification

Image logos used grayscale → color, text logos used muted-gray → brand-color. Same transition timing but visibly different effects. Now both groups use the same grayscale(100%) + opacity 0.55 → grayscale(0%) + opacity 0.9 treatment with shared transition properties. Brand colors (Netflix red, Twilio red, Zurich blue, etc.) sit on the base element and the filter desaturates them by default, mirroring how the image-logos behave.

transform: translateZ(0) on .trust-logo-text to keep inline-SVG + text combos (Zurich, Weather Co, Twilio) from jittering on the filter transition.

Desktop

Fixed:

Mobile

Fixed:

Test plan

• Desktop: load list, scroll, expand a few rows (incl. a Bedrock entry). Confirm thead no longer sticks/overlaps and rows are uniform height.
• Mobile (≤ 768 px): expand a model, verify Pricing/Model Info/Features headings, pricing-card labels and values, info-row stacking, and Python-SDK / Proxy code-snippet sizes all read cleanly without horizontal scroll.
• Mobile expanded panel: confirm code snippet wraps long model names instead of overflowing.
• Trust-logo strip: hover each logo (Stripe, Adobe, Netflix, Zurich, Weather Co, Samsara, etc.) — all should fade between gray and color with the same feel; no positional jitter on Zurich/Weather/Twilio.
• Providers tab: click the A2A provider card → lands on docs/litellm.ai/docs/a2a (200), not the 404.
• Confirm sample_spec no longer appears in the model count or list.