docs(getting_started): standardize on CUDA thread/kernel terminology

[samlaf]

Human Note

Found the getting_started a bit confusing as a gpu learner, as it uses nomenclature that doesn't match cuda's docs. Wondering whether this is intended to make this docs use similar wording as the broader rust-gpu effort, but given that this is rust-cuda, figured being closer to cuda usage of the words would make more sense?

LLM generated

The getting started guide used "invocation" / "multiple invocations of a kernel" to describe parallel execution. That vocabulary comes from the GLSL/SPIR-V/Vulkan compute world — natural for the rust-gpu org, whose flagship project targets SPIR-V, where "invocation" is the official term for a single execution instance of a shader (exactly what CUDA calls a thread). The doc was even internally consistent about it.

The problem is that "invocation" means something different in CUDA-native usage: a "kernel invocation" is the <<<>>> launch — one invocation per launch, which then spawns many threads. So a reader coming from CUDA C++ parses "multiple invocations running in parallel" as multiple grid launches on streams, a separate concept. Since this is rust-cuda, that collision is a real cost.

Standardize on CUDA's own model — kernel = the __global__ function, thread = the unit of parallel execution, launch = the <<<>>> call — and drop "invocation" entirely. This is unambiguous in a CUDA context and easier to follow for beginners.

Also fixes the outright error "mutable state shared by multiple kernels executing in parallel" (there is one kernel; threads execute it), and two incidental typos in the Blocks bullet ("that it execute" → "that execute together", "blocks index avaiable" → "block's index available").