Develop - #72

…PI converters

- Introduced `ConverterCatalog` for fetching and parsing ConvertAPI's OpenAPI spec.
- Added `IConverterCatalog` interface for defining catalog operations.
- Implemented model classes for working with converters and tags.
- Created a `ConverterCatalogDemo` project demonstrating usage of the new service.
- Updated the solution to include the demo project.

#66

Delete the in-SDK scorer (~190 lines: token weighting, format-pair
bonuses, virtual converter synthesis, dead Matches helper). Replace
with a single SearchConvertersAsync that hits the server endpoint and
parses the response with the existing ParseConverters.

API surface:
- New: SearchConvertersAsync(string query)        — async canonical
- New: SearchConverters(string query)             — sync overload
- Kept: SearchConverters(string[] terms)          — joins with spaces

Behavior:
- Empty/whitespace input short-circuits, returns empty list, no
  network call.
- Server orders paths by x-ca-search-rank, so document order is
  relevance order — caller gets best-match-first for free.
- Hard cutover, no local fallback. Older self-hosted CARA servers
  without the ?q= filter will throw InvalidOperationException with
  a clear message.

Smoke-tested against staging:
- "pdf to docx" → 7 results, pdf→docx rank 1
- "pdf to word" → 7 results, pdf→docx rank 1 (was rank 4 with old
  local scorer due to substring bug matching "word" inside
  "password")
- "heic to pdf" → 4 results, heic→pdf rank 1

CHANGELOG entry under 3.1.15-dev (current development cycle).