Add BenchmarkFamily iterable

[nicholasjng]

Designed to consume the arguments of parametrize/product, and yield benchmarks lazily.

Params needs to support lazy iterables as well, to not fall into the eager materialization + memory blowup trap.

---

Testing should contain these scenarios at minimum:

1. Confirm that BenchmarkFamilys are being picked up by nnbench.collect(), similarly to existing collect checks.
2. Add a test that would fail with an eager iterable (e.g. list[Benchmark like it is now), asserting memory consumption stays low e.g. with memray.

Proposal for 2):

Add a benchmark matmul(a: np.ndarray, b: np.ndarray) with two arrays $a, b \in \mathbb{R}^{N \times N}$, parametrize over one of them (five np.random.randns should suffice), assert that memory consumption stays near 3 * N^2 * sizeof(np.float) (a, b, and the intermediary for the result).

[nicholasjng]

Sweet. Results are horrible:

================================================================================= MEMRAY REPORT ==================================================================================
Allocation results for tests/integration/test_benchmark_family_memory_consumption.py::test_foobar at the high watermark

         📦 Total memory allocated: 145.0MiB
         📏 Total allocations: 96
         📊 Histogram of allocation sizes: | █▄  |
         🥇 Biggest allocating functions:
                - matmul:/Users/nicholasjunge/Workspaces/python/nnbench/tests/integration/test_benchmark_family_memory_consumption.py:24 -> 68.1MiB
                - <genexpr>:/Users/nicholasjunge/Workspaces/python/nnbench/tests/integration/test_benchmark_family_memory_consumption.py:22 -> 64.0MiB
                - _call_with_frames_removed:<frozen importlib._bootstrap>:241 -> 267.5KiB

At an allowance of 25MiB, this puts us at 6x the theoretically necessary memory consumption.

[nicholasjng]

Merging, docs updates to follow in a separate PR.