[Performance] ProjectTagViolationDetector reads and parses XML file for every reference × rule combination

## Description

`ProjectTagViolationDetector.GetViolationsFrom()` calls `this.projectTagProvider.GetProjectTag(referenceContext.Reference)` inside a nested loop (for each rule × each reference). `ProjectTagProvider.GetProjectTag()` loads and parses the XML project file from disk on every call.

## Affected Files

- `src/ReferenceCop/Detectors/ProjectTagViolationDetector.cs` (line ~30, inside nested loop)
- `src/ReferenceCop/Providers/ProjectTagProvider.cs` (entire `GetProjectTag` method — `XDocument.Load()` per call)

## Impact

For a solution with R rules and N project references:
- **R × N file I/O operations** and XML parses, even though the same project files are read repeatedly
- For a typical solution (e.g., 20 rules, 50 references), this means 1,000 XML file parses per build, most of them redundant
- Build-time analyzer, so this directly impacts developer inner loop

## Suggested Optimization

Cache the project tag per file path:

```csharp
public class ProjectTagProvider : IProjectTagProvider
{
private readonly ConcurrentDictionary _cache = new();

public string GetProjectTag(string projectFilePath)
{
return _cache.GetOrAdd(projectFilePath, path =>
{
if (!File.Exists(path)) return UnknownProjectTag;
var projectFile = XDocument.Load(path);
return projectFile
.Descendants(PropertyGroupNode)
.Elements(ProjectTagNode)
.FirstOrDefault()?.Value ?? UnknownProjectTag;
});
}
}
```

This reduces N×R file reads to at most N (one per unique project file), typically much fewer.

Start with GetViolationsFrom() in src/ReferenceCop/Detectors/ProjectTagViolationDetector.cs and GetProjectTag() in src/ReferenceCop/Providers/ProjectTagProvider.cs. Confirm how often each project path is requested, then add per-path caching while preserving the existing missing-file and unknown-tag behavior. Done means repeated references avoid redundant XML loads and parsing without changing detected violations.

Written by the indexing model from the issue text.