Advertise real device capabilities in the ESCL server

[zomars]

This closes the TODO in ScanServer.MakeEsclDeviceConfig ("Ideally we want to get the actual device capabilities (flatbed/feeder, resolution etc.)").

What changes

• EsclDeviceConfig gets an optional CapabilitiesProvider callback that can asynchronously supply full EsclCapabilities.
• ScanServer implements the provider by calling ScanController.GetCaps() (the same API the profile dialog uses via DeviceCapsCache) and mapping ScanCaps/PerSourceCaps to the existing EsclCapabilities.PlatenCaps/AdfSimplexCaps/AdfDuplexCaps fields, which so far were only used by the ESCL client.
• EsclApiController.GetScannerCapabilities generates the Platen/Adf sections from those fields when they are populated, advertising only the paper sources the device actually has, with its real resolutions, color modes and max scan area.

Design notes

• Lazy + cached: capabilities are queried on the first ScannerCapabilities request rather than at registration, so sharing devices doesn't open device sessions at startup and server startup is never blocked. A successful query is cached for the lifetime of the registration; a failed query falls back to the defaults and is retried on the next request (same policy as DeviceCapsCache).
• Unchanged fallback: if the driver doesn't provide capabilities (e.g. the current WIA/TWAIN paths), the device is offline, or the query fails, the generated XML is exactly the same as before (Platen + AdfSimplex + AdfDuplex with the default 100-4800 resolution list and A3 max size). No behavior change for setups without caps.
• Resolutions are taken from DpiCaps.CommonValues rather than Values, since some drivers (e.g. Apple/ICA) report a continuous range with thousands of values that would be unreasonable to enumerate in the XML.
• The ESCL client is untouched; standard clients (AirScan/iOS, sane-airscan) that respect the advertised capabilities are the main beneficiaries, since they will no longer offer a non-existent ADF or out-of-range resolutions/page sizes.

Testing

• New ScanServerTests.GetCaps round-trip test (server announces caps from a mock backend, ESCL client parses them back) and ScanServerTests.GetCapsWhenUnavailable fallback test; ran the Remoting test suite plus NAPS2.Escl.Tests on macOS (arm64). One pre-existing failure (ScanWithCorrectOptions, DeviceBusyException when running the whole suite in parallel) reproduces identically on clean master on this machine and is unrelated.
• Tested against real hardware: an EPSON L3210 (flatbed-only) shared through the Apple/ICA driver on macOS. The server now advertises Platen only (no phantom ADF), the real 8.5x11.7" scan area (2550x3509 in 1/300"), the device serial number, and the driver's preferred resolutions (75-9600), and falls back to the previous XML while the device query hasn't succeeded yet. Verified the cached path returns the same document.

Note: for correct results with the Apple driver this depends on #887 (flatbed/feeder swap in DeviceOperator.GetCaps); without it a flatbed-only Mac device is reported as feeder-only. The two PRs are independent code-wise.