fix: refine AI model connection testing

[wsp1911]

Summary

• Limit AI configuration connection checks to a dedicated 5-attempt retry budget while keeping normal AI streaming requests at 10 attempts.
• Expand transport error details and surface localized TLS/certificate, proxy, and network troubleshooting guidance in model connection test results.
• Add API URL validation so model provider URLs must start with http:// or https://.
• Avoid retesting every model after provider saves; only auto-test models affected by connection or request-behavior changes.

Fixes #

Type and Areas

Type:

bug fix / UI/UX

Areas:

AI adapters, web UI, i18n

Motivation / Impact

Connection tests were slow and unclear in SSL/proxy failure cases, often ending with a generic error sending request for url message. This makes connection tests faster, shows more actionable diagnostics, and avoids unnecessary API calls when editing provider model lists.

Users also get immediate validation when the API URL is missing a required HTTP(S) scheme.

Verification

• pnpm run fmt:rs passed
• cargo check -p bitfun-ai-adapters passed
• pnpm run type-check:web passed
• pnpm run i18n:audit passed

Reviewer Notes

• Normal AI streaming retry behavior remains unchanged at 10 attempts.
• The reduced 5-attempt budget only applies to model connection checks.
• Automatic post-save testing now scopes to changed/new models unless provider-level connection settings changed, in which case the saved provider group is tested.

Checklist

• This PR is focused and does not include secrets, temporary prompts, generated scratch files, or unrelated artifacts.
• Relevant verification is recorded above, or skipped checks are explained.
• User-facing strings, docs, and locales are updated where applicable.