perf: cache UTM inverse transforms, MAVLink message info and parameter metadata lookups

[userepo]

Three small lookup fixes, one commit each, smallest first. Each replaces linear or per-call work with a cached or indexed lookup and returns the same answers as before, verified by tests that compare the new code against a copy of the old lookup.

1. utmpos.ToLLA inverse transform cache (ExtLibs/Utilities/utmpos.cs)

ToLLA built a projected coordinate system and a transformation for every point; Grid.CreateGrid calls it once per waypoint on every survey slider change, on the UI thread. The inverse transform is now cached per signed zone, mirroring the forward cache in PointLatLngAlt. Zones outside the real UTM range are converted without caching, so values typed into the coordinate editor cannot grow the cache.

2. GetMessageInfo dictionary lookup (ExtLibs/Mavlink/MavlinkUtil.cs)

GetMessageInfo scanned the 350-entry message table linearly, two to three times per parsed packet, so the cost grew with the message id. The global MAVLINK_MESSAGE_INFOS table is now indexed in a dictionary that is rebuilt when the table is replaced (NvModemMavlinkDialect.Register and plugins do this at runtime); any other array keeps the linear scan. First-match, unknown-id and null-array behavior are unchanged.

Residual: an entry edited in place (MAVLINK_MESSAGE_INFOS[i] = ...) is not seen until the array is replaced. Nothing in this repo or in upstream's example plugin does that.

3. Parameter metadata document indexes (ExtLibs/Utilities/ParameterMetaDataRepositoryAPMpdef.cs, ParameterMetaDataRepositoryAPM.cs)

Every metadata lookup walked the XML linearly: all param elements of up to three pdef sources (local overlay plus downloaded document each), then XContainer.Element over thousands of siblings in the legacy file. Empty answers, the common case for Range, Increment, Bitmask and RebootRequired, are never cached, so filling a parameter table cost about 1.4 s every time, partly on the UI thread.

Each pdef document is now indexed by name attribute, keeping document order so scoped and unscoped entries resolve exactly as before; the legacy document is indexed by vehicle and parameter element name. An index is tied to its document and rebuilt when the document is replaced or edited. Empty answers are still not cached, because the metadata downloads asynchronously after first use.

Pre-existing and left alone: a vehicle type with no .apm.pdef.xml on disk still costs a File.Exists per lookup, and _parameterMetaDataXML is an unsynchronized dictionary written from a task continuation.

Measurements

Release build, .NET 10, Windows x64.

Workload	Before	After
100,000 utmpos.ToLLA calls	197-221 ms	16-26 ms
Grid.CreateGrid, 20,519 points	49 ms	5 ms
10 million GetMessageInfo calls, mixed ids	1,863 ms	146-155 ms
Tlog parse, 2M packets of common low-id messages	627-655 ns/packet	568-585 ns/packet
Tlog parse, 2M packets of ArduPilot-specific messages	1,073-1,096 ns/packet	571-575 ns/packet
Parameter metadata, 1,200 names x 8 keys, first pass	1,714-1,744 ms	27 ms
Same, warm passes	1,371-1,468 ms	9-14 ms

Verification

• New tests: UtmposTransformCacheTests (11), MavlinkMessageInfoLookupTests (8), ParameterMetadataIndexTests (10), over the packaged metadata files and synthetic documents that cover the ordering rules.
• 451,176 parameter metadata answers through the public ParameterMetaDataRepository (6,836 names, 11 keys, 6 vehicle types including the SITL, AP_Periph and legacy fallbacks) are byte-identical between old and new code.
• Full suite green apart from the tests that already fail on Windows on master.