Add Olmo3 implementation

[2015aroras]

This PR adds the upcoming Olmo 3. The main architectural differences from Olmo 2 are:

• Sliding window attention is used for 3 out of 4 layers. RoPE scaling is not applied to sliding window attention layers.

Since the architecture is very similar to Olmo 2, this PR opts to merge Olmo 3 changes into the Olmo 2 implementation (similar to vllm-project/vllm#24534). I can create a separate Olmo 3 implementation instead if preferred.

[2015aroras]

I used the model conversion example for testing. I got the following results when using bf16 on shanearora/2025-sep-a-base-model, modified to have yarn rope scaling enabled.

📈 METRICS
==============================
MSE (Mean Squared Error):     1.592396e-02
Reference Variance:           6.831117e+00
NMSE:                         2.331092e-03
Max Absolute Error:           0.438750
Mean Absolute Error:          0.116665
NMSE (dB):                    -26.32 dB

🎯 INTERPRETATION
==============================
👍 Good match

📋 GUIDANCE
==============================
👍 GOOD: Your GGML conversion is working well.
   Small differences are likely due to precision/quantization.

📚 NMSE BENCHMARKS
==============================
✅ RESULT: PASS (NMSE = 2.33e-03)

Also, below is the allenai/OLMo-2-0425-1B with fp32.

📈 METRICS
==============================
MSE (Mean Squared Error):     1.594746e-03
Reference Variance:           9.219801e+00
NMSE:                         1.729697e-04
Max Absolute Error:           0.168732
Mean Absolute Error:          0.033951
NMSE (dB):                    -37.62 dB

🎯 INTERPRETATION
==============================
👍 Very good match

📋 GUIDANCE
==============================
✅ EXCELLENT: Your GGML conversion is working very well!
   The differences are negligible for practical use.

📚 NMSE BENCHMARKS
==============================
✅ RESULT: PASS (NMSE = 1.73e-04)

[CISC]

Suggested change

	if (is_swa) {
	// For sliding window layers, Olmo3 does not use rope scaling.
	// This is achieved here by setting freq_scale and attn_factor to 1.
	// We also set ext_factor to 0 to avoid a few unnecessary computations.
	Qcur = ggml_rope_ext(
	ctx0, Qcur, inp_pos, nullptr,
	n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
	0.0, 1.0, beta_fast, beta_slow
	);
	Kcur = ggml_rope_ext(
	ctx0, Kcur, inp_pos, nullptr,
	n_rot, rope_type, n_ctx_orig, freq_base, 1.0,
	0.0, 1.0, beta_fast, beta_slow
	);
	}
	else {
	if (!is_swa) {

[2015aroras]

This if block is needed. For SWA layers Olmo2 uses standard rope. Removing that if block would remove rope on SWA layers.

[2015aroras]

6997fad Clarified comment slightly