Make RNG seed optional and improve per-core seed generation for bernoulli/uniform/rand

[suzusuzu]

Ticket

Problem description

ttnn.bernoulli, ttnn.uniform, and ttnn.rand currently take a uint32_t seed with 0 acting as a "no-seed / random" sentinel.
This has two issues:

1. The API cannot express "no seed" explicitly, and seed=0 is ambiguous (it could be a valid user seed).
2. The device program previously derived per-core seeds by incrementing the base seed. In practice, this can lead to highly correlated/random streams across adjacent seeds (e.g., results for seed=1 and seed=2 end up looking almost identical), reducing randomness quality and surprising users expecting independent outputs.
   • ref:

     tt-metal/ttnn/cpp/ttnn/operations/rand/device/rand_program_factory.cpp

     Lines 109 to 110 in 0839f47

     	// Each core has its own seed to increase the number of generated random numbers
     	uint32_t seed = operation_attributes.seed != 0 ? operation_attributes.seed + i : get_random_seed();

What's changed

• Updated the APIs to use std::optional<uint32_t> for seeds:
  • Python bindings now default seed=None instead of 0.
  • Operation attributes store std::optional<uint32_t> rather than a uint32_t with a sentinel.
• Updated program hash computation to ignore seeds by setting the cached seed to std::nullopt (consistent with prior behavior where the seed should not affect compilation/caching).
• Reworked per-core seed generation:
  • Instead of incrementing the base seed, we now initialize an std::mt19937 once per program creation/override using either the user-provided seed or std::time(nullptr) when no seed is provided.
  • Each core receives a fresh seed sampled from a uniform distribution, improving independence between cores and avoiding the "seed=1 vs seed=2 are almost the same" behavior.

Checklist

• 
• 
• 
• New/Existing tests provide coverage for changes

Model tests

If your changes cover model-related code, you should run tests corresponding to affected models and platforms (Single card, T3K, Galaxy). "Choose your pipeline" workflows facilitate running multiple kinds of tests in a single run. Each offers models-mandatory and models-extended presets.
The former includes a minimal set of tests, to be run always. The latter extends that with additional ones - use your best judgement in deciding which is the most appropriate for your PR.

• 

  • models-mandatory preset (runs: Device perf regressions and Frequent model and ttnn tests)
  • models-extended preset (runs: the mandatory tests, plus Demo and Model perf tests)
  • other selection - specify runs

• 

  • models-mandatory preset (runs: Unit tests)
  • models-extended preset (runs: the mandatory tests, plus Demo and Model perf tests)
  • other selection - specify runs

• 

  • models-mandatory preset (runs: Quick tests)
  • models-extended preset (runs: the mandatory tests, plus Demo and Model perf tests)
  • other selection - specify runs

[suzusuzu]

I verified the behavior of the updated bernoulli / uniform / rand operations as follows:

>>> ttnn.rand([3, 3], device=device)
ttnn.Tensor([[ 0.2637,  0.6328,  0.0532],
             [ 0.8672,  0.9336,  0.4629],
             [ 0.6523,  0.8281,  0.6133]], shape=Shape([3, 3]), dtype=DataType::BFLOAT16, layout=Layout::TILE)

>>> input = ttnn.to_device(ttnn.from_torch(torch.empty(3, 3, dtype=torch.bfloat16).uniform_(0, 1), layout=ttnn.TILE_LAYOUT), dev
ice=device)
>>> ttnn.bernoulli(input)
ttnn.Tensor([[ 0.0000,  1.0000,  1.0000],
             [ 1.0000,  0.0000,  1.0000],
             [ 0.0000,  0.0000,  0.0000]], shape=Shape([3, 3]), dtype=DataType::FLOAT32, layout=Layout::TILE)

>>> input = ttnn.to_device(ttnn.from_torch(torch.ones(3, 3, dtype=torch.bfloat16), layout=ttnn.TILE_LAYOUT), device=device)
>>> ttnn.uniform(input)
ttnn.Tensor([[ 0.9102,  0.9531,  0.6406],
             [ 0.4531,  0.2266,  0.8164],
             [ 0.5195,  0.7578,  0.0889]], shape=Shape([3, 3]), dtype=DataType::BFLOAT16, layout=Layout::TILE)