Fact-check and correct migrated Antora docs

[michalharakal]

Fact-check pass over the Antora docs that landed in #494 / #495 + polish from #497. The conversion from markdown preserved every stale claim the source files had accumulated since whenever they were last touched; this issue tracks the real inaccuracies surfaced on review.

Confirmed findings

1. Stale version numbers (8 references)

Docs reference `0.13.0` (Java guides) and `0.5.0` (io-readers). Current release is `0.18.0`, next is `0.19.0`.

• `how-to/java-cli-app.adoc` : 2 refs
• `how-to/java-llm-inference.adoc` : 1 ref
• `how-to/java-model-training.adoc` : 1 ref
• `how-to/io-readers.adoc` : 2 refs (at `0.5.0`)
• `tutorials/java-getting-started.adoc` : 2 refs

2. Wrong Maven group (2 references)

`io-readers.adoc` lines 23 and 33 use `sk.ainet.core:skainet-io-gguf` / `sk.ainet.core:skainet-io-onnx`. Actual group is `sk.ainet` (no `.core`), confirmed in root `build.gradle.kts` line 17 and `skainet-bom/build.gradle.kts` line 6.

3. References to code that moved to SKaiNET-transformers (49 references across 3 pages)

Per the 2026-04-13 repo split, LLM runtimes live in the sibling `SKaiNET-transformers` repo. Mainline SKaiNET's engine layer has no `KLlamaJava`, `KBertJava`, `JavaAgentLoop`, or any `sk.ainet.apps.kllama.` / `sk.ainet.apps.bert.` / `sk.ainet.apps.kllama.chat.*` package. The following three pages describe APIs that no longer live here:

• `tutorials/kllama-getting-started.adoc` (entire page)
• `how-to/java-cli-app.adoc` (entire page — it's a KLlama CLI tutorial)
• `how-to/java-llm-inference.adoc` (entire page — KLlama + KBert + agent loop)

4. Pandoc link artifact

`explanation/perf/jvm-cpu.adoc` line 28 still uses `link:java-25-cpu-backend.md[...]` — wrong extension, wrong Antora syntax. Should be `xref:explanation/perf/java-25-cpu-backend.adoc[...]`.

5. JDK prerequisite inconsistency

• `java-getting-started`, `java-cli-app`, `java-llm-inference`, `java-model-training` say `JDK 21 or later`
• `explanation/perf/jvm-cpu` says `JDK 21+ (JDK 22 toolchain)`
• `how-to/build` says `Sets up JDK 25`
• CI actually runs JDK 25 in `docs.yml`, `build.yml`, `publish.yml`

Not wrong per se — `JDK 21+` is the true minimum — but inconsistent. Normalize to `JDK 21+ (CI uses JDK 25)`.

Fix plan (one PR, several commits)

1. Bump stale version numbers to `0.19.0` (the upcoming release — see CHANGELOG `[Unreleased]` section). Covers Add GGUFReader #1.
2. Fix Maven group in `io-readers.adoc`. Covers Add suppor for running MNIST sample app #2.
3. Replace the three moved-LLM pages with short redirect stubs pointing at `SKaiNET-transformers`. Remove them from the nav. Covers MNIST data set loader #3.
4. Fix the pandoc link artifact and normalize JDK prereq wording. Covers convolution layer (Conv2D) #4 and Dropout #5.

Each commit verified with a clean local Antora build (zero warnings, zero errors).