feat(llama-cpp): bump llama.cpp and add moe_cache_mib option (#12729)

Bump LLAMA_VERSION to b76b13379d3b (upstream 0.7.0 plus later
commits). This is the same one-line change as the bot PR #12725.

Add the moe_cache_mib model option. It sets upstream's
moe_cache_size, a GPU cache for MoE experts kept in host memory,
from a size in MiB. It is off by default. Invalid, negative or
non-numeric values are ignored with a log line. Forks that predate
the field ignore the option, using the same prepare.sh probe as the
other fork-skew switches.

Document the option, and cover the MiB to bytes conversion with a
standalone unit test.

Assisted-by: Claude:claude-sonnet-5-5 g++

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>
Co-authored-by: Ettore Di Giacinto <mudler@localai.io>

backend/cpp/llama-cpp/moe_cache_option_test.cpp

@@ -0,0 +1,36 @@
+// SPDX-License-Identifier: MIT
+
+#include "moe_cache_option.h"
+
+#include <cstdio>
+
+static int failures = 0;
+
+static void expect(const std::string & in, std::optional<size_t> want) {
+    const auto got = llama_grpc::parse_moe_cache_bytes(in);
+    if (got != want) {
+        std::fprintf(stderr, "parse_moe_cache_bytes(\"%s\") returned a wrong result\n", in.c_str());
+        failures++;
+    }
+}
+
+int main() {
+    // MiB to bytes.
+    expect("0", 0);
+    expect("1", 1024ull * 1024ull);
+    expect("2048", 2048ull * 1024ull * 1024ull);
+    // 8 GiB must not overflow a 32-bit intermediate.
+    expect("8192", 8192ull * 1024ull * 1024ull);
+
+    // Invalid values are rejected so the caller keeps the default.
+    expect("-1", std::nullopt);
+    expect("", std::nullopt);
+    expect("abc", std::nullopt);
+    expect("12abc", std::nullopt);
+    expect("1.5", std::nullopt);
+    expect("99999999999999999999999", std::nullopt);
+    // Too large to express in bytes.
+    expect("9223372036854775807", std::nullopt);
+
+    return failures == 0 ? 0 : 1;
+}