fix: partial descriptor behavior

[inemtsev]

Fixes #1512

Problem

PluginGeneratedSerialDescriptor instances created without a generated serializer throw
ArrayIndexOutOfBoundsException from hashCode(), equals() and toString(). This happens
for serializers produced by the @Serializer(forClass = ...) companion shortcut:

@Serializable(Model.Companion::class)
data class Model(val a: Int, val b: String) {
    @Serializer(forClass = Model::class)
    companion object
}

For these, the plugin constructs the descriptor with generatedSerializer = null, so
childSerializers is empty while elementsCount > 0 and all three identity methods walk
getElementDescriptor(i), which indexes into the empty array.

The bug is invisible on ordinary encode/decode (nothing calls hashCode), but it breaks any
feature that uses descriptors as cache keys in DescriptorSchemaCache. This is exactly the
crash reported in #1512 (open since 2021). The stack trace there is this path verbatim:
DescriptorSchemaCache.get -> PluginGeneratedSerialDescriptor.hashCode -> AIOOBE. The
originally reported trigger is useAlternativeNames = true (the documented limitation:
JsonCustomSerializersTest must run with useAlternativeNames = false), and a later report
in that thread hits the same crash via coerceInputValues = true. Because the root cause is
descriptor identity itself, every schema-cache entry point is affected the same way: in #3189
it forced a try-catch around the schema-cache lookup, and any future descriptor-keyed cache
would inherit the same landmine.

Fix

Identity of such "incomplete" descriptors is now computed from data the descriptor itself
owns: serial name, type parameters, kind, and element names, instead of the child descriptors
it cannot provide:

• Complete descriptors are untouched: same hash values, same equality semantics as before.
• Incomplete descriptors hash/compare by serialName + typeParams + kind + element names.
• Complete vs incomplete never compare equal (previously threw), which is what keeps the
  equals/hashCode contract intact with two different hash formulas. The kind check also
  prevents accidental equality with EnumDescriptor, which reuses the null-serializer
  construction but overrides hashCode itself.
• toString() for incomplete descriptors prints element names without child type names.

The new branch executes only inside the existing lazy _hashCode initializer and in equals,
so there is no hot-path cost (verified: TwitterFeedBenchmark decode/encode at parity with
master).

Benchmark	master (ops/s)	this PR (ops/s)	Δ
decodeTwitter	1591 ± 22	1596 ± 38	+0.3%, intervals overlap (parity)
encodeTwitter	3244 ± 12	3200 ± 54	-1.4%, intervals overlap (parity)

[inemtsev]

Hi @sandwwraith @fzhinkin, please take a look at this improvement.

It fixes a crash path (that was commented for @JsonNames annotation in the past), but also allows me to clean up try/catch in #3189

Would be great to apply this fix first before cleaning up the above PR.