feat(stjm): add per-target undiscriminated source migration support

[egil]

Why

When adopting Egil.SystemTextJson.Migration on an existing type, stored JSON typically has no $type discriminator. Today, discriminator-less object payloads are deserialized directly as the target type, which makes it hard to incrementally introduce a renamed/restructured target with a migration from the prior shape.

Approach

Adds an opt-in JsonMigratableAttribute.UndiscriminatedSourceType that lets a target declare a single source CLR type to assume when an object payload has no recognized discriminator. The converter then runs the matching migrator (static IMigrateFrom<> on the target, or an external IMigrate<,> registration). Behavior without the attribute is unchanged.

public record CustomerNameV0(string FirstName, string LastName);

[JsonMigratable(
    TypeDiscriminator = "customer-name-v1",
    UndiscriminatedSourceType = typeof(CustomerNameV0))]
public record CustomerNameV1(string Name) : IMigrateFrom<CustomerNameV0, CustomerNameV1>
{
    public static bool TryMigrateFrom(CustomerNameV0 source, out CustomerNameV1 result)
    {
        result = new CustomerNameV1($"{source.FirstName} {source.LastName}");
        return true;
    }
}

Internals:

• TypeMetadata carries the configured undiscriminated source type.
• JsonMigratableConverterFactory resolves the matching MigratorReference once per converter and throws a clear InvalidOperationException at setup time if no migrator matches.
• JsonMigratableConverter.Inspect consults the resolved migrator at the empty-object branch and after the source-discriminator scan, before the dictionary-payload fallback.

Performance

Added BenchmarkDotNet coverage for the new undiscriminated source migration route with a plain System.Text.Json manual migration baseline. Refreshed the generated docs/perf reports and README performance summary from those results. Ran the full perf suite on this machine in both directions: changed branch first, then a detached main worktree baseline.

Common benchmark rows compared: 72. New changed-branch benchmark rows: 12. No allocation deltas were observed across common rows. The only common-row slowdowns were small/noisy: +5.2% on a plain STJ serialize baseline at tag count 2, and +0.6% to +0.9% on existing external migration rows; all other common rows were equal or faster in this run.

New undiscriminated source migration benchmark results:

• Reflection: 555.0 ns / 1,032 B at 2 tags; 1,554.2 ns / 2,992 B at 32 tags; 9,609.7 ns / 17,400 B at 256 tags.
• SourceGen: 587.8 ns / 1,008 B at 2 tags; 1,847.2 ns / 2,968 B at 32 tags; 8,540.4 ns / 17,376 B at 256 tags.

Notes for reviewers

• Discriminator matches still take priority. The opt-in only activates when no discriminator is recognized, so existing payloads that include $type (or a known source discriminator) keep their current routing.
• Only one source type is supported per target by design, to keep the choice deterministic. A test on UndiscriminatedStaticTarget declares two IMigrateFrom<> impls and verifies the configured one wins.
• Misconfiguration (no migrator for the declared source) fails at converter creation with both type names in the message, rather than silently passing through.
• Docs (README, recipes) and the LegacyPayloads sample are updated with the new pattern; mdsnippets refreshed.
• Validation: 184 library tests, 44 sample tests, perf project Release build, and full BenchmarkDotNet suite on changed branch plus main baseline.